from atomgit_hub import snapshot_download
snapshot_download("Ascend-SACT/tacotron2",  local_dir = './download')

# 1. 加载昇腾 CANN 环境变量
source /usr/local/Ascend/ascend-toolkit/set_env.sh

# 2. 安装基础依赖 (使用清华源加速)
# 必须锁定 torch==2.1.0 和 torch-npu==2.1.0.post6
pip install torch==2.1.0 torch-npu==2.1.0.post6 -i https://pypi.tuna.tsinghua.edu.cn/simple

# 3. 安装关键音频库 (核心步骤：严格匹配版本)
pip install torchaudio==2.1.0 -i https://pypi.tuna.tsinghua.edu.cn/simple

# 4. 安装应用层依赖
# sentencepiece 是 SpeechT5 分词器必需的
pip install transformers datasets soundfile aria2 accelerate sentencepiece

# 1. 创建目录
mkdir -p /root/autodl-tmp/tacotron2-persian
cd /root/autodl-tmp/tacotron2-persian

# 2. 多线程下载 (16线程)
aria2c -x 16 -s 16 -c https://hf-mirror.com/MahtaFetrat/Persian-Tacotron2-on-ManaTTS/resolve/main/synthesizer.pt

# 3. 文件名修复 (若 aria2 保存为 hash 文件名)
# 如果 ls 看到文件名是一串乱码，执行下面这行；如果是 synthesizer.pt 则跳过
# mv e0494ea356ff... synthesizer.pt

import torch
import torch_npu
import sys

MODEL_PATH = "synthesizer.pt"
DEVICE = torch.device("npu:0")

print(f"🔧 环境检查: PyTorch {torch.__version__}")
if not torch.npu.is_available():
    print("❌ NPU 不可用")
    sys.exit(1)

print(f"📂 加载权重: {MODEL_PATH}")
try:
    # 1. CPU 读取
    checkpoint = torch.load(MODEL_PATH, map_location='cpu')
    print("✅ CPU 读取成功")

    # 2. 提取状态字典
    model_state = checkpoint.get('model_state', checkpoint)
    
    # 3. NPU 写入测试
    print("🔥 开始 NPU 显存写入测试...")
    success = 0
    for i, (k, v) in enumerate(model_state.items()):
        if i >= 20: break # 只测前20个以节省时间
        if isinstance(v, torch.Tensor):
            _ = v.to(DEVICE) # 关键动作：迁移至 NPU
            sys.stdout.write(".")
            success += 1
            
    print(f"\n🎉 成功迁移 {success} 个张量至 NPU")
    print("✅ 结论: Tacotron 2 权重格式与昇腾 NPU 兼容。")

except Exception as e:
    print(f"❌ 失败: {e}")

python check_and_load.py

🎉 成功迁移 20 个张量至 NPU

✅ 结论: Tacotron 2 权重格式与昇腾 NPU 兼容

import torch
import torch_npu
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
from datasets import load_dataset
import soundfile as sf
import time

# 配置
DEVICE = torch.device("npu:0")
OUTPUT_FILE = "npu_generated_audio.wav"

def main():
    print(f"🚀 初始化 NPU TTS... 设备: {DEVICE}")

    # 1. 加载模型 (自动下载)
    processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
    model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(DEVICE)
    vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(DEVICE)

    # 2. 处理输入
    text = "Successful verification! The Huawei Ascend NPU is generating this audio efficiently."
    inputs = processor(text=text, return_tensors="pt")
    inputs = {k: v.to(DEVICE) for k, v in inputs.items()} # 迁移至 NPU

    # 3. 加载声纹 (随机替代)
    # 由于 HuggingFace 数据集脚本废弃，使用随机向量进行功能验证
    speaker_embeddings = torch.randn(1, 512).to(DEVICE)

    # 4. 推理
    print("🔥 开始生成...")
    start = time.time()
    with torch.no_grad():
        speech = model.generate_speech(inputs["input_ids"], speaker_embeddings, vocoder=vocoder)
    cost = time.time() - start

    # 5. 保存
    sf.write(OUTPUT_FILE, speech.cpu().numpy(), samplerate=16000)
    print(f"⏱️ 耗时: {cost:.4f}s | 🎉 音频已保存: {OUTPUT_FILE}")

if __name__ == "__main__":
    main()

HF_ENDPOINT=https://hf-mirror.com python npu_tts_gen.py

⏱️ 耗时: 2.1415s | 🎉 音频已保存: npu_generated_audio.wav

错误现象	根本原因	解决方案
ImportError: libsentencepiece	缺少分词库	pip install sentencepiece
Torchaudio not found	环境缺失或版本不匹配	严禁直接 pip install。必须执行 pip install torchaudio==2.1.0
Network is unreachable	连接 HuggingFace 官网超时	运行时添加 HF_ENDPOINT=https://hf-mirror.com
aria2 文件名为乱码	HTTP Header 解析问题	手动 mv 重命名文件为 .pt 后缀
UserWarning: AutoNonVariable...	NPU 底层算子提示	属于正常日志，忽略即可