MMS-300M-1130-Forced-Aligner 昇腾NPU适配文档

python >= 3.9
torch == 2.9.0+cpu
torch_npu == 2.9.0.post1
transformers >= 4.40.0
soundfile >= 0.12
numpy >= 1.24

pip install torch==2.9.0+cpu --index-url https://download.pytorch.org/whl/cpu
pip install torch_npu==2.9.0.post1  # 根据昇腾版本选择
pip install transformers soundfile numpy

# NPU 推理（自动选择NPU）
python inference.py \
  --model_path . \
  --audio sample.wav \
  --transcript "hello world" \
  --device npu \
  --output alignment_result.json

# CPU 推理（用于对比验证）
python inference.py \
  --model_path . \
  --audio sample.wav \
  --transcript "hello world" \
  --device cpu \
  --output alignment_result_cpu.json

# 同时评测 CPU 和 NPU
python evaluate.py \
  --model_path . \
  --audio sample.wav \
  --devices cpu,npu \
  --warmup 2 \
  --runs 5 \
  --output eval_result.json

os.environ["PYTORCH_NPU_ALLOC_CONF"] = "expandable_segments:True"

import torch
import torch_npu
from inference import run_inference

# 自动选择 NPU
result = run_inference(
    model_path="./mms-300m-1130-forced-aligner",
    audio_path="sample.wav",
    transcript="hello world this is a test",
    device_type="npu",  # 或 "cpu" / "cuda"
    dtype=torch.float32,
)

print(f"推理设备: {result['device']}")
print(f"推理耗时: {result['infer_time']:.3f}s")
print(f"RTF: {result['rtf']:.4f}")

for seg in result['segments']:
    print(f"{seg['token']:12s} {seg['start_time']:7.3f}s - {seg['end_time']:7.3f}s")

============================================================
Running inference on: npu
============================================================
[INFO] Loading model from: .
[INFO] Target device: npu, dtype: torch.float32
[INFO] Running in eager mode (torch.compile disabled for stability)
[TIME] Model load: 1.72s
[TIME] Audio load: 0.00s (duration: 5.00s)
[TIME] Inference: 0.22s (RTF: 0.044)
[TIME] Alignment: 3.21s

[RESULT] Aligned 41 segments

[INFO] Results saved to: alignment_result_npu.json

[ALIGNMENT RESULTS]
  h                0.000s -   0.020s
  e                0.020s -   0.040s
  l                0.040s -   0.060s
  l                0.060s -   0.080s
  o                0.080s -   0.100s
  ...