MusicGen-Medium 昇腾 NPU 部署文档

pip install torch torch_npu --index-url https://pypi.ascend.huawei.com/simple
pip install transformers scipy soundfile

/opt/atomgit/musicgen-medium-npu/
├── inference.py              # 推理脚本（主入口）
├── model_cache/              # 模型权重 (~7.5GB)
│   ├── config.json
│   ├── generation_config.json
│   ├── pytorch_model.bin.*   # 分布式权重文件
│   └── ...
├── output/                   # 输出目录
└── README.md                 # 本文件

cd /opt/atomgit/musicgen-medium-npu

# 生成 8 秒音乐
python3 inference.py \
  --prompt "lo-fi music with a soothing melody" \
  --duration 8 \
  --seed 42

# 使用自定义参数
python3 inference.py \
  --prompt "energetic electronic dance music" \
  --duration 10 \
  --guidance_scale 3.0 \
  --temperature 1.0 \
  --top_k 50 \
  --output output/electronic.wav

cd /opt/atomgit/musicgen-medium-npu

python3 inference.py \
  --prompt "calm piano music with soft strings" \
  --duration 4 \
  --max_new_tokens 200 \
  --guidance_scale 3.0 \
  --seed 42 \
  --output output/test_output.wav

============================================================
MusicGen-Medium 昇腾 NPU 推理
============================================================
Prompt: lo-fi music with a soothing melody
Duration: 4.0s
Max new tokens: 512
Guidance scale: 3.0
Do sample: True
Temperature: 1.0
Top-k: 50, Top-p: 0.95
Seed: 42
============================================================
[INFO] Loading model from: /opt/atomgit/musicgen-medium-npu/model_cache
[INFO] Using NPU device: 0
[INFO] NPU name: Ascend910_9362
[INFO] NPU memory: 61.3 GB
[INFO] Loading processor...
[INFO] Loading model...

Config of the text_encoder: T5EncoderModel
  T5Config: d_model=768, d_ff=3072, num_heads=12, num_layers=12, vocab_size=32128

Config of the audio_encoder: EncodecModel
  EncodecConfig: hidden_size=128, codebook_size=2048, sampling_rate=32000

Config of the decoder: MusicgenForCausalLM
  MusicgenDecoderConfig: hidden_size=1536, ffn_dim=6144,
    num_attention_heads=24, num_hidden_layers=48, num_codebooks=4, vocab_size=2048

[INFO] Model loaded in 34.6s
[INFO] Model parameters: 2.01B
[INFO] Audio sampling rate: 32000 Hz
[INFO] Adjusted max_new_tokens to: 200
[INFO] Generating music for prompt: 'lo-fi music with a soothing melody'
[INFO] Generation parameters: {'max_new_tokens': 200, 'do_sample': True,
       'guidance_scale': 3.0, 'temperature': 1.0, 'top_k': 50, 'top_p': 0.95}

[INFO] Generation completed in 16.62s
[INFO] Audio shape: torch.Size([1, 1, 126080])
[INFO] Audio saved to: output/log_test.wav
[INFO] Audio duration: 3.94s
[INFO] Audio samples: 126080

============================================================
性能统计
============================================================
生成耗时: 16.62s
生成速度: 12.0 tokens/s
音频时长: 3.94s
NPU 显存 - Allocated: 3.75GB, Reserved: 4.43GB
============================================================

python3 -c "
import soundfile as sf
import numpy as np
data, sr = sf.read('output/test_output.wav')
print(f'Duration: {len(data)/sr:.2f}s')
print(f'Sample rate: {sr} Hz')
print(f'RMS: {np.sqrt(np.mean(data**2)):.4f}')
print(f'Has NaN: {np.isnan(data).any()}')
"

{
  "test": "NPU reproducibility",
  "seed": 42,
  "max_tokens": 150,
  "run1_time_s": 14.6,
  "run2_time_s": 5.5,
  "cosine_similarity": 0.9999999999981538,
  "outputs_identical": true,
  "both_no_nan": true,
  "both_no_inf": true,
  "conclusion": "PASS"
}