Reader v2 (6 layers, LibriSpeech 960 h) replaces reader v1 on macOS, iOS and the h19p compiles
76865ae verified | { | |
| "what": "the app's self-test with reader v2 against the r4 bars (r8: fixture >= 19/20, long WER <= torch + 0.1 pt)", | |
| "device": "mac", | |
| "form": "gpu", | |
| "inputs": { | |
| "gate": "<work>/exports/app/mac_v2/gate-gpu-114102/result.json", | |
| "long": "<work>/exports/app/mac_v2/long-gpu-114107/result.json", | |
| "iso": "<work>/exports/app/extractor_isolation_mac_v1_gpu_124853.json", | |
| "transcribe": [] | |
| }, | |
| "references": { | |
| "ref_fixtures": { | |
| "file": "ref_fixtures_coreai_path_v2.json", | |
| "finished": "2026-10-15T11:40:04+0900", | |
| "transcripts_same_as_torch": 20 | |
| }, | |
| "torch": "exports/reader_v2/ref_T750.npz, exports/reader_v2/fp16/hyp_gpu_T750.tsv (v2)" | |
| }, | |
| "device_record": { | |
| "arch": "h16c", | |
| "available_mb": null, | |
| "footprint_mb": 3.392088, | |
| "machine": "Mac16,9", | |
| "os": "Version 27.0 (Build 26A428)", | |
| "os_build": "26A428", | |
| "physical_memory_gb": 137.44, | |
| "power": { | |
| "battery": "mac", | |
| "level": null, | |
| "low_power_mode": false | |
| }, | |
| "thermal": "nominal" | |
| }, | |
| "model": { | |
| "cache": { | |
| "cleared": false, | |
| "fingerprint": "96edfd3630bf84a5cfbf7e703d4e0b7c40ddb5724ff92a723430a1e1646806de", | |
| "mode": "check", | |
| "ms": 0.774125, | |
| "previous": "96edfd3630bf84a5cfbf7e703d4e0b7c40ddb5724ff92a723430a1e1646806de" | |
| }, | |
| "folder": "eg2-transcribe", | |
| "form": "gpu", | |
| "platform_folders": [ | |
| "macos" | |
| ], | |
| "reader_config_version": "v2", | |
| "stages": [ | |
| { | |
| "aot": false, | |
| "bundle": "eg2_audio_fp16_T3000.aimodel", | |
| "compute": "gpu", | |
| "load_s": 0.810113875, | |
| "name": "encoder", | |
| "warm_up_ms": 107.578167 | |
| }, | |
| { | |
| "aot": false, | |
| "bundle": "eg2_reader_v2_fp16_T750.aimodel", | |
| "compute": "gpu", | |
| "load_s": 1.140881792, | |
| "name": "reader", | |
| "warm_up_ms": 101.645459 | |
| } | |
| ], | |
| "vocabulary": 200000, | |
| "window": { | |
| "drop_last": 1, | |
| "frames": 3000, | |
| "left_overlap": 96, | |
| "tokens": 750 | |
| } | |
| }, | |
| "fixtures": { | |
| "source": "gate-gpu-114102", | |
| "utterances": 20, | |
| "features_max_abs_diff": 4.76837158203125e-07, | |
| "fp32_encoder_cos_min_app_features": 0.9999999995026775, | |
| "fp32_encoder_cos_min_hf_features": 0.9999999990170414, | |
| "iso_features_max_abs": 4.76837158203125e-07, | |
| "extractor_pass": true, | |
| "transcripts_same_as_torch": 20, | |
| "mismatches": [], | |
| "fixture_pass": true, | |
| "fp16_soft_cos_mean_app_features": 0.9999941175021174, | |
| "fp16_soft_cos_min_app_features": 0.989784733563025, | |
| "fp16_soft_cos_min_reference_features": 0.9964469062042872, | |
| "labels_same_both_features": 20, | |
| "host_labels_equal_python_gpu_path_on_features_npy": 20, | |
| "host_fixture_pass": true | |
| }, | |
| "long": { | |
| "source": "long-gpu-114107", | |
| "audio_s": 3257.44325, | |
| "decode_s": 0.644676, | |
| "log_mel_s": 0.050381916, | |
| "log_mel_cpu_s": { | |
| "blocks": 637, | |
| "fft": 0.490767789, | |
| "log": 0.07565346800000003, | |
| "mel": 0.12388825099999985, | |
| "note": "CPU seconds summed over the parallel blocks of 512 frames: framing + FFT + |X|, mel product, log" | |
| }, | |
| "torch_wer": { | |
| "wer": 0.03812316715542522, | |
| "errors": 377, | |
| "ref_words": 9889, | |
| "hyp_words": 9931, | |
| "engine": "word Levenshtein (numpy); jiwer: bench_pipeline.py score" | |
| }, | |
| "python_path": { | |
| "sequential_s": 8.9276, | |
| "overlap_s": 7.4047, | |
| "overlap_s_pass2": 7.4072, | |
| "errors": 378, | |
| "ref_words": 9889, | |
| "wer": 0.03822428961472343, | |
| "x_realtime_overlap": 439.9, | |
| "x_realtime_sequential": 364.9, | |
| "reader_bundle": "eg2_reader_v2_fp16_T750_gpu_h16c.aimodelc", | |
| "json": [ | |
| "pipeline_long_fp16_gpu_overlap.json", | |
| "pipeline_long_fp16_gpu_overlap.json" | |
| ], | |
| "labels": "labels_long_fp16_gpu_overlap.npy" | |
| }, | |
| "features_max_abs_diff_long": 4.76837158203125e-07, | |
| "host_long_labels_differing_on_features_npy": 0, | |
| "host_long_word_distance_on_features_npy": 0, | |
| "host_long_pass": true, | |
| "modes": { | |
| "overlap_off": { | |
| "wall_s": [ | |
| 8.055587666, | |
| 8.048949625 | |
| ], | |
| "x_realtime_windows": [ | |
| 404.37065364561823, | |
| 404.7041417531545 | |
| ], | |
| "x_realtime_end_to_end": 372.25176353851333, | |
| "stage_ms_median": [ | |
| 40.072459, | |
| 23.918042 | |
| ], | |
| "stage_ms_median_every_pass": [ | |
| [ | |
| 40.072459, | |
| 23.918042 | |
| ], | |
| [ | |
| 40.055458, | |
| 23.883417 | |
| ] | |
| ], | |
| "host_ms_per_window": { | |
| "collapse": 0.002702672, | |
| "feed": 0.315168056, | |
| "inputs": 0.035913679999999996, | |
| "labels": 0.008363376000000004, | |
| "total": 0.362147784 | |
| }, | |
| "wer": { | |
| "errors": 379, | |
| "reference_words": 9889, | |
| "wer": 0.03832541207402164 | |
| }, | |
| "wer_errors_every_pass": [ | |
| 379, | |
| 379 | |
| ], | |
| "word_distance_to_python_hyp": 2, | |
| "labels_differing_from_python": 10, | |
| "labels_differing_positions_s": [ | |
| 559.84, | |
| 907.2, | |
| 1238.8, | |
| 1389.28, | |
| 1472.48, | |
| 1678.92, | |
| 2213.44, | |
| 2592.36, | |
| 2673.44, | |
| 2688.28 | |
| ], | |
| "top1_logit_max": 67.125, | |
| "labels_differing_pass2_vs_pass1": 0, | |
| "soft_tokens_first_minutes_vs_continuous": { | |
| "cos_mean": 0.9999893888428746, | |
| "cos_min": 0.9470607199456169, | |
| "reference": "soft_tokens_continuous.npy", | |
| "tokens": 15000 | |
| } | |
| }, | |
| "overlap_on": { | |
| "wall_s": [ | |
| 7.403074375, | |
| 8.219681333 | |
| ], | |
| "x_realtime_windows": [ | |
| 440.01222802789954, | |
| 396.2979972133671 | |
| ], | |
| "x_realtime_end_to_end": 402.2462381381712, | |
| "stage_ms_median": [ | |
| 58.493667, | |
| 33.972166 | |
| ], | |
| "stage_ms_median_every_pass": [ | |
| [ | |
| 58.493667, | |
| 33.972166 | |
| ], | |
| [ | |
| 58.649375, | |
| 33.906042 | |
| ] | |
| ], | |
| "host_ms_per_window": { | |
| "collapse": 0.005405664, | |
| "feed": 0.2686709280000001, | |
| "inputs": 0.03527999200000001, | |
| "labels": 0.010198040000000002, | |
| "total": 0.3195546240000001 | |
| }, | |
| "wer": { | |
| "errors": 379, | |
| "reference_words": 9889, | |
| "wer": 0.03832541207402164 | |
| }, | |
| "wer_errors_every_pass": [ | |
| 379, | |
| 379 | |
| ], | |
| "word_distance_to_python_hyp": 2, | |
| "labels_differing_from_python": 10, | |
| "labels_differing_positions_s": [ | |
| 559.84, | |
| 907.2, | |
| 1238.8, | |
| 1389.28, | |
| 1472.48, | |
| 1678.92, | |
| 2213.44, | |
| 2592.36, | |
| 2673.44, | |
| 2688.28 | |
| ], | |
| "top1_logit_max": 67.125, | |
| "labels_differing_pass2_vs_pass1": 0 | |
| } | |
| }, | |
| "errors": [ | |
| 379, | |
| 379, | |
| 379, | |
| 379 | |
| ], | |
| "long_vs_torch_pass": true, | |
| "long_vs_python_pass": true, | |
| "overlap_labels_same_as_sequential": true, | |
| "speed": { | |
| "python_sequential_s": 8.9276, | |
| "python_overlap_s": 7.4047, | |
| "app_sequential_s": 8.055587666, | |
| "app_overlap_s": 7.403074375, | |
| "sequential_ratio": 0.9023239914422688, | |
| "overlap_ratio": 0.9997804603832701 | |
| }, | |
| "speed_pass": true | |
| }, | |
| "transcribe": [], | |
| "verdicts": { | |
| "extractor_pass": true, | |
| "fixture_pass": true, | |
| "host_fixture_pass": true, | |
| "host_long_pass": true, | |
| "long_vs_torch_pass": true, | |
| "long_vs_python_pass": true, | |
| "speed_pass": true | |
| }, | |
| "pass": true, | |
| "finished": "2026-10-15T11:41:51+0900" | |
| } |