MOSS-SoundEffect (昇腾 NPU 适配版) #NPU

# 安装依赖
pip install torch==2.9.0
pip install torch_npu==2.9.0.post1
pip install transformers>=4.57.0
pip install torchaudio
pip install modelscope

# 下载 MOSS-SoundEffect 模型
modelscope download --model openmoss/MOSS-SoundEffect --local_dir ./MOSS-SoundEffect

# 下载 MOSS-Audio-Tokenizer (音频编解码器)
modelscope download --model openmoss/MOSS-Audio-Tokenizer --local_dir ./MOSS-Audio-Tokenizer

import torch
import torch_npu
import torchaudio
from pathlib import Path
from transformers import AutoModel, AutoProcessor

# 使用 NPU
DEVICE = "npu" if torch.npu.is_available() else "cpu"
DTYPE = torch.bfloat16

# 加载 Processor (指定 codec_path 为本地音频 Tokenizer 路径)
processor = AutoProcessor.from_pretrained(
    "MOSS-SoundEffect",
    trust_remote_code=True,
    codec_path="MOSS-Audio-Tokenizer",
)
processor.audio_tokenizer = processor.audio_tokenizer.to(DEVICE)

# 模型加载 (~70s)
model = AutoModel.from_pretrained(
    "MOSS-SoundEffect",
    trust_remote_code=True,
    attn_implementation="sdpa",  # NPU 使用 SDPA 后端
    torch_dtype=DTYPE,
).to(DEVICE)
model.eval()

# 生成音效
prompt = "雷声隆隆，雨声淅沥。"
conversations = [[processor.build_user_message(ambient_sound=prompt)]]
batch = processor(conversations, mode="generation")
input_ids = batch["input_ids"].to(DEVICE)
attention_mask = batch["attention_mask"].to(DEVICE)

with torch.no_grad():
    outputs = model.generate(
        input_ids=input_ids,
        attention_mask=attention_mask,
        max_new_tokens=4096,
    )

# 保存音频
for i, message in enumerate(processor.decode(outputs)):
    audio = message.audio_codes_list[0]
    torchaudio.save(f"output_{i}.wav", audio.unsqueeze(0), 24000)

python inference_npu.py

python evaluate_accuracy.py

python benchmark_perf.py

@misc{MOSS-TTS,
  author = {OpenMOSS Team},
  title = {MOSS-TTS Family},
  year = {2026},
  publisher = {MOSI.AI},
}