mlboydaisuke's picture
Reader v2 (6 layers, LibriSpeech 960 h) replaces reader v1 on macOS, iOS and the h19p compiles
76865ae verified
Raw History Blame Contribute Delete
6.32 kB
{
"form": "fp16",
"tokens": 750,
"window_s": 30.0,
"name": "eg2_reader_v2_fp16_T750",
"name_tag": "v2",
"head": "argmax",
"io_dtype": "float16",
"reader": {
"kind": "release",
"weights": "<work>/exports/reader_v2/reader/reader.pt",
"weights_sha256": "d605f6fba8687497be6b5fd717ea6c51beaee5d5f69c30a64225da96ebe0d8c2",
"arch": "a",
"k": null,
"layers": 6,
"params": 19573762,
"scale": 216.3509979248047,
"n_vocab": 200000,
"table": "<work>/exports/reader_v2/reader/word_table.npy",
"reader_dir": "<work>/exports/reader_v2/reader",
"model_py_sha256": "d71d2bac4cf14f98f91029978c67ace2d940b6f01e1213350cfee330bfdbefe9",
"decode_py_sha256": "b447a21b5031f201ba60353e45c6cb13b32e97a9294f4c61728e1d8e29d35b2f",
"config": {
"version": "v2",
"window": {
"seconds": 30,
"left_overlap_tokens": 96,
"right_extra_tokens": 1,
"note": "encoder is causal; the 83 tokens at the start of a window and its last token differ from a continuous run (fixtures/long-speckled-band: soft_tokens_w30 vs soft_tokens_continuous)"
},
"decode": "greedy argmax over N+1, collapse repeats, drop blank",
"scale": 216.3509979248047,
"wer": {
"dev-clean": {
"wer": 0.03754,
"utterances": 2703,
"ref_words": 54402
},
"dev-other": {
"wer": 0.10263,
"utterances": 2864,
"ref_words": 50948
},
"test-clean": {
"wer": 0.04207,
"utterances": 2620,
"ref_words": 52576
},
"test-other": {
"wer": 0.10422,
"utterances": 2939,
"ref_words": 52343
},
"fixtures": {
"wer": 0.03697,
"utterances": 20,
"ref_words": 541
},
"decode": "greedy, no language model"
}
}
},
"norms": "max-abs rescaled, LayerNorm written out (float16)",
"software": {
"python": "3.11.13",
"torch": "2.9.0",
"coreai-core": "1.0.0b2",
"coreai-torch": "0.4.1",
"coreai-opt": "0.2.1"
},
"inputs": {
"soft_tokens": [
1,
750,
512
],
"token_mask": [
1,
750
]
},
"outputs": {
"top1_index": [
1,
750
],
"top1_logit": [
1,
750
],
"blank_logit": [
1,
750
]
},
"eager_gate": {
"utterances": 2,
"transcripts_same": 2,
"token_agreement": 1.0,
"tokens": 349,
"max_abs_top1_logit": 0.1364002227783203,
"max_abs_blank_logit": 0.06196784973144531,
"rows": [
{
"set": "fixture",
"id": "1089-134686-0000",
"tokens": 261,
"tokens_same": 261,
"transcript_same": true,
"max_abs_top1_logit": 0.10126399993896484,
"max_abs_blank_logit": 0.06196784973144531
},
{
"set": "test-clean",
"id": "6930-75918-0000",
"tokens": 88,
"tokens_same": 88,
"transcript_same": true,
"max_abs_top1_logit": 0.1364002227783203,
"max_abs_blank_logit": 0.040701866149902344
}
]
},
"decomposed_gate": {
"utterances": 2,
"transcripts_same": 2,
"token_agreement": 1.0,
"tokens": 349,
"max_abs_top1_logit": 0.1364002227783203,
"max_abs_blank_logit": 0.06196784973144531,
"rows": [
{
"set": "fixture",
"id": "1089-134686-0000",
"tokens": 261,
"tokens_same": 261,
"transcript_same": true,
"max_abs_top1_logit": 0.10126399993896484,
"max_abs_blank_logit": 0.06196784973144531
},
{
"set": "test-clean",
"id": "6930-75918-0000",
"tokens": 88,
"tokens_same": 88,
"transcript_same": true,
"max_abs_top1_logit": 0.1364002227783203,
"max_abs_blank_logit": 0.040701866149902344
}
]
},
"convert_s": 0.2,
"optimize_s": 0.0,
"op_counts": {
"builtin.module": 1,
"coreai.abs": 15,
"coreai.argmax": 1,
"coreai.cast": 1,
"coreai.concat": 12,
"coreai.constant": 115,
"coreai.decomposable.broadcasting_add": 80,
"coreai.decomposable.broadcasting_batch_matmul": 30,
"coreai.decomposable.broadcasting_maximum": 15,
"coreai.decomposable.broadcasting_mul": 141,
"coreai.decomposable.broadcasting_sub": 15,
"coreai.gelu": 6,
"coreai.graph": 2,
"coreai.invoke": 6,
"coreai.output": 2,
"coreai.reduce_max": 16,
"coreai.reduce_mean": 28,
"coreai.reduce_sum": 1,
"coreai.reshape": 28,
"coreai.rsqrt": 30,
"coreai.slice": 42,
"coreai.softmax": 1,
"coreai.transpose": 51
},
"save_s": 0.4,
"bundle": "<work>/exports/reader_v2/fp16/eg2_reader_v2_fp16_T750.aimodel",
"bytes": 346661157,
"main_hash": "5f7948f5f6e1de11191cdffcdcc4c9ae9c6ef012a786c782488c1e766fa60c81",
"coreai_build": "coreai-build 3600.83.1",
"inspect_source": {
"returncode": 0,
"json": "<work>/exports/reader_v2/fp16/inspect_eg2_reader_v2_fp16_T750.json"
},
"aot": {
"gpu": {
"argv": "compile <work>/exports/reader_v2/fp16/eg2_reader_v2_fp16_T750.aimodel --output <work>/exports/reader_v2/fp16/eg2_reader_v2_fp16_T750_gpu_h16c.aimodelc --platform macOS --preferred-compute gpu --architecture h16c",
"returncode": 0,
"seconds": 3.3,
"log_tail": [
"Compiling h16c: 1 of 1"
],
"ane_validation_messages": 0,
"bundle": "<work>/exports/reader_v2/fp16/eg2_reader_v2_fp16_T750_gpu_h16c.aimodelc",
"bytes": 346749479,
"ane_regions": 0,
"ane_region_names": [],
"files": [
"main-h16c-delegates",
"main-h16c.mlirb",
"main.hash",
"metadata.json",
"stats.json"
],
"inspect": {
"returncode": 0,
"json": "<work>/exports/reader_v2/fp16/inspect_eg2_reader_v2_fp16_T750_gpu_h16c.json"
}
},
"ane": {
"argv": "compile <work>/exports/reader_v2/fp16/eg2_reader_v2_fp16_T750.aimodel --output <work>/exports/reader_v2/fp16/eg2_reader_v2_fp16_T750_ane_h16c.aimodelc --platform macOS --preferred-compute neural-engine --architecture h16c",
"returncode": 0,
"seconds": 4.0,
"log_tail": [
"Compiling h16c: 1 of 1"
],
"ane_validation_messages": 0,
"bundle": "<work>/exports/reader_v2/fp16/eg2_reader_v2_fp16_T750_ane_h16c.aimodelc",
"bytes": 347380273,
"ane_regions": 2,
"ane_region_names": [
"main_9wemocz7YmEwKSHhsD5JpQ_0_ANE_region_0",
"main_9wemocz7YmEwKSHhsD5JpQ_0_ANE_region_1"
],
"files": [
"main-h16c-delegates",
"main-h16c.mlirb",
"main.hash",
"metadata.json",
"stats.json"
],
"inspect": {
"returncode": 0,
"json": "<work>/exports/reader_v2/fp16/inspect_eg2_reader_v2_fp16_T750_ane_h16c.json"
}
}
},
"seconds": 158.6,
"finished": "2026-10-15T10:17:15+0900"
}