Commit History
-
ExecuTorch: delegate the sv/seg LoRA adapters to XNNPACK
Verified signatureThe speaker heads' rank-16/32 adapters were traced with their weights reshaped in the graph, so XNNPACK refused every adapter conv and they ran on the portable kernel. The export now gives them conv-shaped constant weights. seg_8000ms is ~10x faster per 8 s window at one worker (small int8 1593 -> 129 ms, medium int8 1360 -> 140 ms; sv 1.5 s 132 -> 11 ms), and now scales with threads. Same weights and size; fp32 output unchanged; ops.txt drops aten::convolution.out. DER re-checked: voxconverse 12 small 12.5 / medium 12.1 (was 12.8 / 12.0), aishell5 all 18 int8 / fp32 small 33.0 / 32.9, medium 32.4 / 32.1.a030328
-
metadata: sv.normalised -> sv.normalized (consistent spelling)
Verified signatureb3b0236 -
PCEC keep input: refreshed reference scripts (feed 0 by default), self-contained modeling file, merges in ONNX metadata
Verified signaturef3bf13b
-
metadata: BPE merges beside the vocab (tokenizer buildable from metadata.json)
Verified signature865e2d7 -
vad stream on the model's frame grid: 320-sample tail, score i of call c = frame 100c + i - 4
Verified signatureab59913
-
vad method carries its state (7 in / 7 out): exact one-call-per-second FSMN VAD
Verified signatured1ce15e -
v2: per-dataset benchmarks, duration-named methods, schema v1
Verified signaturefeaad3b -
asr4all+ m interim: card rewritten (transformers-first, per-dataset tables pending), embed_speaker / segment_speakers / diarize.py on the transformers surface, class labels, public metadata without build paths
Verified signature2b68981 -
asr4all+ m: full GPTQ coverage on the speaker heads, sv rung ladder, FSMN VAD on all surfaces, Diarizer-first card with through-.pte EER/DER (rel-2026-09-18)
Verified signature216bd23