Higgs-Audio-V2 昇腾 NPU 部署文档

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

/opt/atomgit/higgs-audio-npu/
├── inference.py              # 推理脚本（主入口）
├── download.sh               # 模型下载脚本
├── convert_weights.py         # 权重重映射工具
├── eval_audio.py              # 精度/性能评测脚本
├── README.md                  # 本文件
├── model_cache/               # 生成模型权重 (~11GB)
│   ├── model-00001-of-00003.safetensors
│   ├── model-00002-of-00003.safetensors
│   ├── model-00003-of-00003.safetensors
│   ├── config.json
│   └── ...
└── tokenizer_cache/           # 音频 tokenizer 权重 (~11GB)
    ├── model.safetensors
    └── config.json

cd /opt/atomgit/higgs-audio-npu
bash download.sh

cd /opt/atomgit/higgs-audio-npu

# 使用推荐种子 (seed=1, 已验证效果)
python3 inference.py --text "In the quiet of the morning, birds begin to sing as the first rays of sunlight paint the sky in shades of orange and pink." --seed 1

# 使用简短文本
python3 inference.py --text "Hello, welcome to the world of audio generation." --seed 37

# 中文文本
python3 inference.py --text "清晨的阳光下，小鸟在枝头欢快地歌唱。" --seed 42

# 推荐：长描述性文本 + seed=1，生成约 81.6s 高质量音频
cd /opt/atomgit/higgs-audio-npu
python3 inference.py \
  --text "In the quiet of the morning, birds begin to sing as the first rays of sunlight paint the sky in shades of orange and pink. The gentle breeze carries the sweet scent of blooming flowers across the meadow, while a distant brook murmurs its timeless melody." \
  --seed 1

# 简短文本
python3 inference.py --text "Hello, this is the Higgs Audio generation system running on Ascend NPU." --seed 37

使用设备: npu:0
Higgs-Audio-V2 昇腾 NPU 推理
============================================================
模型参数量: 5.38B
输入文本: In the quiet of the morning, birds begin to sing...
生成参数: max_new_tokens=2048, temperature=0.3, top_k=50, top_p=0.95
============================================================
输入: [1, 49] tokens
生成完成, 耗时: 84.10s, 生成 2048 tokens, 速度: 24.4 tokens/s
音频序列: torch.Size([1, 2048, 8])
有效音频帧: 2048
音频解码: 0.35s
输出保存: output_audio/output_1.wav

# 运行全量评测
python3 eval_audio.py --mode all

# 仅精度评测
python3 eval_audio.py --mode precision --seeds 1 37 42 74

# 仅性能评测
python3 eval_audio.py --mode performance --tokens 512 1024 2048