VibeVoice-1.5B Ascend NPU 适配

# 基础依赖
pip install torch==2.9.0 torch_npu transformers>=4.51.0

# VibeVoice 模型包
pip install vibevoice

# 音频处理
pip install torchaudio soundfile scipy

# 模型下载
pip install modelscope
modelscope download --model microsoft/VibeVoice-1.5B

import torch
import torch_npu
from vibevoice.processor import VibeVoiceProcessor
from vibevoice.modular.modeling_vibevoice_inference import VibeVoiceForConditionalGenerationInference

# 加载模型
model_path = "/path/to/VibeVoice-1___5B"
qwen_tokenizer_path = "/path/to/Qwen2___5-1___5B"  # Qwen2.5-1.5B tokenizer
processor = VibeVoiceProcessor.from_pretrained(
    model_path, language_model_pretrained_name=qwen_tokenizer_path
)
model = VibeVoiceForConditionalGenerationInference.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
)

# 移到 NPU
device = torch.device("npu:0")
model = model.to(device)
model.eval()

# 生成播客音频
script = (
    "Speaker 1: Welcome to our podcast. Today we explore the latest in AI technology.\n"
    "Speaker 2: It's exciting to see how speech synthesis has advanced in recent years."
)
inputs = processor(text=script, return_tensors="pt", padding=True)
device_inputs = {k: v.to(device) if isinstance(v, torch.Tensor) else v
                 for k, v in inputs.items()}

with torch.no_grad():
    output = model.generate(**device_inputs, max_new_tokens=512,
                           do_sample=True, return_speech=True)

# 保存音频
import torchaudio
audio = output.speech_outputs[0].cpu()
torchaudio.save("output.wav", audio, 24000)

# 基础推理 (默认演示脚本)
python inference.py

# 指定文本推理
python inference.py --text "Speaker 1: Hello from Ascend NPU.\nSpeaker 2: The model is running well."

# 运行所有评测 (精度 + 性能)
python inference.py --all

# 仅精度评测
python eval_accuracy.py

# 仅性能测试
python eval_performance.py

Speaker 1: <说话人1的文本>
Speaker 2: <说话人2的文本>
Speaker 1: <后续对话>

import os
import torch_npu

# 内存分配器优化
os.environ["PYTORCH_NPU_ALLOC_CONF"] = "expandable_segments:True"

# 设备初始化
device = torch.device("npu:0")
torch.npu.set_device(device)

# 同步操作 (确保执行完成)
torch.npu.synchronize()

VibeVoice-1.5B/
├── inference.py              # 主推理脚本
├── eval_accuracy.py          # 精度评测脚本
├── eval_performance.py       # 性能基准测试
├── README.md                 # 本文档
├── output/                   # 输出目录
│   ├── results.json          # 运行结果
│   ├── accuracy_*.json       # 精度评测结果
│   ├── performance_*.json    # 性能测试结果
│   └── demo_*.wav            # 演示音频输出
└── model/                    # 模型文件 (需下载)

@misc{microsoft2025vibevoice,
  title={VibeVoice: A Frontier Open-Source Text-to-Speech Model},
  author={Microsoft Research},
  year={2025},
  howpublished={\url{https://huggingface.co/microsoft/VibeVoice-1.5B}},
}