MOSS-VoiceGenerator 昇腾 NPU 部署文档

# 核心依赖
torch >= 2.9.0
transformers >= 5.0.0
torch_npu >= 2.9.0
torchaudio
modelscope

# Python 版本
Python >= 3.9

# 安装 transformers 5.x
pip install "transformers>=5.0.0"

# 安装 modelscope
pip install modelscope

# 安装 torch_npu (昇腾驱动环境)
pip install torch_npu

# 主模型
modelscope download --model openmoss/MOSS-VoiceGenerator

# Audio Tokenizer (必需)
modelscope download --model openmoss/MOSS-Audio-Tokenizer

cd /opt/atomgit/moss-voicegenerator-npu
python inference.py

from modelscope import AutoModel, AutoProcessor
import torch

# 加载模型
model = AutoModel.from_pretrained(
    "openmoss/MOSS-VoiceGenerator",
    trust_remote_code=True,
    attn_implementation="sdpa",  # NPU 使用 sdpa
    torch_dtype=torch.bfloat16,
).to("npu")

processor = AutoProcessor.from_pretrained(
    "openmoss/MOSS-VoiceGenerator",
    trust_remote_code=True,
)

# 准备输入
conversations = [[processor.build_user_message(
    text="你好，世界。",
    instruction="温柔的年轻女性声音。"
)]]

# 推理
batch = processor(conversations, mode="generation")
outputs = model.generate(
    input_ids=batch["input_ids"].to("npu"),
    attention_mask=batch["attention_mask"].to("npu"),
)

# 解码
decoded = processor.decode(outputs)
audio = decoded[0].audio_codes_list[0]

/opt/atomgit/moss-voicegenerator-npu/
├── inference.py                    # 推理脚本
├── npu_compat.py                   # NPU 兼容层
├── test_npu_modelscope.py          # 完整测试脚本
├── test_precision_deterministic.py # 精度评测脚本
├── readme.md                       # 本文档
├── adaptation_report.md           # 适配验证报告
└── output/                         # 生成音频目录
    ├── chinese_female.wav
    └── english_male.wav