File size: 6,322 Bytes
bd15125 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 | from __future__ import annotations
import argparse
import gc
import json
from pathlib import Path
import mlx.core as mx
import numpy as np
import torch
from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoProcessor, AutoTokenizer
from ark_asr_mlx import ArkASR
from ark_asr_mlx.conversion import DEFAULT_REVISION, DEFAULT_SOURCE
def _cosine_similarity(left: np.ndarray, right: np.ndarray) -> float:
left_flat = left.astype(np.float64).reshape(-1)
right_flat = right.astype(np.float64).reshape(-1)
denominator = np.linalg.norm(left_flat) * np.linalg.norm(right_flat)
return float(np.dot(left_flat, right_flat) / denominator)
def _source_outputs(source_path: Path, audio_path: Path) -> dict[str, object]:
device = "mps" if torch.backends.mps.is_available() else "cpu"
dtype = torch.bfloat16 if device == "mps" else torch.float32
processor = AutoProcessor.from_pretrained(source_path, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(
source_path,
trust_remote_code=True,
fix_mistral_regex=True,
)
model = AutoModelForCausalLM.from_pretrained(
source_path,
trust_remote_code=True,
dtype=dtype,
attn_implementation="sdpa",
).to(device)
model.eval()
conversation = [
{
"role": "user",
"content": [
{"type": "audio", "path": str(audio_path)},
{"type": "text", "text": "Please transcribe this audio."},
],
}
]
inputs = processor.apply_chat_template(
conversation,
add_generation_prompt=True,
return_tensors="pt",
sampling_rate=16_000,
audio_padding="longest",
text_kwargs={"padding": "longest"},
audio_max_length=30 * 16_000,
)
raw_input_features = inputs["audios"].float().numpy()
inputs = inputs.to(device)
inputs["audios"] = inputs["audios"].to(dtype=dtype)
eos_ids = tokenizer.eos_token_id
keep_ids = {eos_ids} if isinstance(eos_ids, int) else set(eos_ids or [])
bad_ids = set(tokenizer.all_special_ids) - keep_ids
bad_ids.update(
token_id
for token, token_id in tokenizer.get_added_vocab().items()
if token.startswith("<") and token.endswith(">") and token_id not in keep_ids
)
bad_words_ids = [[token_id] for token_id in sorted(bad_ids)]
with torch.inference_mode():
adapted = model.audio_encoder(inputs["audios"])
outputs = model(**inputs, use_cache=False)
generated = model.generate(
**inputs,
do_sample=False,
max_new_tokens=256,
pad_token_id=tokenizer.pad_token_id,
eos_token_id=tokenizer.eos_token_id,
bad_words_ids=bad_words_ids,
)
generated_ids = generated[:, inputs.input_ids.shape[1] :]
text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
result = {
"input_ids": inputs.input_ids.detach().cpu().numpy(),
"input_features": raw_input_features,
"adapted": adapted.float().detach().cpu().numpy(),
"last_logits": outputs.logits[:, -1, :].float().detach().cpu().numpy(),
"token_ids": generated_ids.detach().cpu().numpy()[0].tolist(),
"text": text,
}
del generated, outputs, adapted, inputs, model, processor, tokenizer
gc.collect()
if device == "mps":
torch.mps.empty_cache()
return result
def validate(model_path: Path, audio_path: Path) -> dict[str, object]:
source_path = Path(
snapshot_download(repo_id=DEFAULT_SOURCE, revision=DEFAULT_REVISION)
)
source = _source_outputs(source_path, audio_path)
asr = ArkASR.from_pretrained(model_path)
processed = asr.processor.process(audio_path)
mlx_ids = mx.array(processed.input_ids)
mlx_features = mx.array(processed.input_features).astype(
asr.model.audio_encoder.whisper.conv1.weight.dtype
)
adapted = asr.model.audio_encoder(mlx_features)
embeddings = asr.model.prepare_prompt_embeddings(
mlx_ids,
mlx_features,
processed.audio_start,
processed.audio_count,
)
logits = asr.model(mlx_ids, input_embeddings=embeddings)
mx.eval(adapted, logits)
mlx_result = asr.transcribe(audio_path)
source_tokens = list(source["token_ids"])
if source_tokens and source_tokens[-1] == asr.model.config.eos_token_id:
source_tokens.pop()
metrics = {
"input_ids_equal": np.array_equal(processed.input_ids, source["input_ids"]),
"input_features_max_abs": float(
np.max(
np.abs(
np.array(mlx_features.astype(mx.float32))
- source["input_features"]
)
)
),
"adapter_cosine": _cosine_similarity(
np.array(adapted.astype(mx.float32)),
source["adapted"],
),
"last_logits_cosine": _cosine_similarity(
np.array(logits[:, -1, :].astype(mx.float32)),
source["last_logits"],
),
"token_ids_equal": list(mlx_result.token_ids) == source_tokens,
"text_equal": mlx_result.text == source["text"],
"mlx_text": mlx_result.text,
"pytorch_text": source["text"],
}
if not metrics["input_ids_equal"]:
raise AssertionError(f"Input IDs differ: {metrics}")
if metrics["input_features_max_abs"] > 1e-6:
raise AssertionError(f"Input features differ: {metrics}")
if metrics["adapter_cosine"] < 0.999:
raise AssertionError(f"Adapter parity failed: {metrics}")
if metrics["last_logits_cosine"] < 0.999:
raise AssertionError(f"Logit parity failed: {metrics}")
if not metrics["token_ids_equal"] or not metrics["text_equal"]:
raise AssertionError(f"Generation parity failed: {metrics}")
return metrics
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("audio", type=Path)
parser.add_argument(
"--model",
type=Path,
default=Path("models/ARK-ASR-0.6B-bf16"),
)
args = parser.parse_args()
print(json.dumps(validate(args.model, args.audio), indent=2, ensure_ascii=False))
if __name__ == "__main__":
main()
|