1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
|
# 请预先运行以下指令安装一些工具
brew install portaudio ffmpeg sox
# 安装 Conda，地址 https://repo.anaconda.com/archive/
Anaconda3-2025.12-2-MacOSX-arm64.pkg
# 创建虚拟环境
conda create -n qwen3-tts python=3.12 -y
# 激活虚拟环境
conda activate qwen3-tts
# 安装系统依赖
pip install -U qwen-tts
# 进入主目录
cd Qwen3-TTS
# 安装 Python 依赖项
pip install -e .
# 验证 MPS
(qwen3-tts) xxxMac-mini Qwen3-TTS % python
Python 3.12.12 | packaged by Anaconda, Inc. | (main, Oct 21 2025, 20:07:49) [Clang 20.1.8 ] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> print(torch.backends.mps.is_available())
True
>>> print(torch.backends.mps.is_built())
True
>>> exit()
# 测试
cd examples
python test_model_12hz_base.py
# 报错：SoX could not be found!
# 安装 SOX
brew install sox
# 无效的路径格式：ERROR：Invalid path format detected
# 删除模型路径 "/"
# 修改 MODEL_PATH = "Qwen/Qwen3-TTS-12Hz-1.7B-Base/" 为 MODEL_PATH = "Qwen/Qwen3-TTS-12Hz-1.7B-Base"
# 测试
python test_model_12hz_base.py
# ERROR：ImportError: FlashAttention2 has been toggled on, but it cannot be used the package flash_attn seems to be not installed
# FlashAttention / FlashAttention2 只支持 NVIDIA CUDA GPU，不支持 Apple Silicon
# 修改支持 Mac M 系列芯片的代码
# tts = Qwen3TTSModel.from_pretrained(
# MODEL_PATH,
# device_map=device,
# dtype=torch.bfloat16,
# attn_implementation="flash_attention_2",
# )
tts = Qwen3TTSModel.from_pretrained(
MODEL_PATH,
torch_dtype=torch.bfloat16, # M4 芯片完美支持 bfloat16
attn_implementation="sdpa", # 改为 sdpa，这是 Mac MPS 硬件加速支持的优化方式
device_map="mps", # 强制使用 Mac M4 的 GPU
)
# 测试
python test_model_12hz_base.py
# ERROR：Torch 不支持编译 CUDA 的方法 （M 芯片没有 CUDA）
# 使用 M 系列芯片的同步指令
def run_case(tts: Qwen3TTSModel, out_dir: str, case_name: str, call_fn):
if torch.cuda.is_available():
torch.cuda.synchronize()
elif torch.backends.mps.is_available():
torch.mps.synchronize() # 这是 Mac M4 对应的同步指令
# torch.cuda.synchronize()
t0 = time.time()
wavs, sr = call_fn()
if torch.cuda.is_available():
torch.cuda.synchronize()
elif torch.backends.mps.is_available():
torch.mps.synchronize() # 这是 Mac M4 对应的同步指令
# torch.cuda.synchronize()
t1 = time.time()
print(f"[{case_name}] time: {t1 - t0:.3f}s, n_wavs={len(wavs)}, sr={sr}")
for i, w in enumerate(wavs):
sf.write(os.path.join(out_dir, f"{case_name}_{i}.wav"), w, sr)
# 测试
python test_model_12hz_base.py
|
piliqiu