[Win] Deployment issue and setup improvement proposal

[nightt5879]

📌 Summary

This issue is to track the setup and deployment process of MegaTTS3 on Windows, including:

• pynini build errors via pip
• Suggested conda workaround
• Version compatibility issues (e.g., gradio, pydantic)
• Directory layout for checkpoints
• Proposal to include Gradio-based WebUI

---

✅ Environment

• OS: Windows 11
• Python: 3.10
• pip: latest
• Conda: yes

---

🔧 Problems Encountered

1. pynini build fails via pip, now i tring the conda install -c conda-forge pynini=2.1.5

[zhangbo0405]

您好，assets文件夹下Chinese_prompt.npy与English_prompt.npy如何根据.wav文件生成的，什么规格的

[Idiotabtcodes]

I've been trying to set up the environment too on Windows 10. You can check here for references.

[nightt5879]

@jiang-ziyue 我这两天在上课，回来看到readme里面有win部分的，我就测试了一下（按照顺序）但是还是有问题

(megatts3-env) E:\robot\project\MegaTTS3>python tts/infer_cli.py --input_wav 'assets/chinese_prompt.wav' --input_text "另一边的桌上,一位读书人嗤之以鼻道,'佛子三藏,神子燕小鱼是什么样的人物,李家的那个李子夜如何与他们相提并论？'" --output_dir ./gen --p_w 2.0 --t_w 3.0
Traceback (most recent call last):
  File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 18, in <module>
    import librosa
ModuleNotFoundError: No module named 'librosa'

我自己有进行另一个方法的尝试，我尽量这段时间试试吧，在提交PR之前我会找一些朋友测试 + 创建fork的分支 届时我们可以进一步讨论

[jiang-ziyue]

@nightt5879 好的谢谢

[nightt5879]

@jiang-ziyue 请问这个问题要如何解决，是我的transformer版本不对吗？ 我尝试了重新下载g2p的资源 但是不太行，搜索 + 问AI也没太解决，想来问一下

(megatts_win) E:\robot\project\MegaTTS3>python tts/infer_cli.py --input_wav 'assets/chinese_prompt.wav' --input_text "你好世界，我是中文测试" --output_dir ./gen --p_w 2.0 --t_w 3.0

A module that was compiled using NumPy 1.x cannot be run in
NumPy 2.1.3 as it may crash. To support both 1.x and 2.x
versions of NumPy, modules must be compiled with NumPy 2.0.
Some module may need to rebuild instead e.g. with 'pybind11>=2.12'.

If you are a user of the module, the easiest solution will be to
downgrade to 'numpy<2' or try to upgrade the affected module.
We expect that some modules will need time to support NumPy 2.

Traceback (most recent call last):  File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 267, in <module>
    infer_ins = MegaTTS3DiTInfer()
  File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 90, in __init__
    self.build_model(self.device)
  File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 112, in build_model
    self.dur_model = ARDurPredictor(
  File "E:\robot\project\MegaTTS3\tts\modules\ar_dur\ar_dur_predictor.py", line 174, in __init__
    super().__init__(hparams, hidden_size, dec_hidden_size, lm_num_layers, dict_size, code_size)
  File "E:\robot\project\MegaTTS3\tts\modules\ar_dur\ar_dur_predictor.py", line 61, in __init__
    self.encoder = FS_ENCODERS[self.hparams['encoder_type']](self.hparams, dict_size)
  File "E:\robot\project\MegaTTS3\tts\modules\ar_dur\ar_dur_predictor.py", line 31, in <lambda>
    'rel_fft': lambda hp, dict_size: RelTransformerEncoder(
  File "E:\robot\project\MegaTTS3\tts\modules\ar_dur\commons\rel_transformer.py", line 366, in __init__
    self.emb = Embedding(n_vocab, hidden_channels, padding_idx=0)
  File "E:\robot\project\MegaTTS3\tts\modules\ar_dur\commons\layers.py", line 60, in Embedding
    m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx)
  File "E:\anaconda\envs\megatts_win\lib\site-packages\torch\nn\modules\sparse.py", line 145, in __init__
    self.reset_parameters()
  File "E:\anaconda\envs\megatts_win\lib\site-packages\torch\nn\modules\sparse.py", line 154, in reset_parameters
    init.normal_(self.weight)
  File "E:\anaconda\envs\megatts_win\lib\site-packages\torch\nn\init.py", line 175, in normal_
    return _no_grad_normal_(tensor, mean, std, generator)
  File "E:\anaconda\envs\megatts_win\lib\site-packages\torch\nn\init.py", line 20, in _no_grad_normal_
    return tensor.normal_(mean, std, generator=generator)
E:\anaconda\envs\megatts_win\lib\site-packages\torch\nn\init.py:20: UserWarning: Failed to initialize NumPy: _ARRAY_API not found (Triggered internally at ..\torch\csrc\utils\tensor_numpy.cpp:84.)
  return tensor.normal_(mean, std, generator=generator)
| loaded 'dur_model' from './checkpoints\duration_lm/model_only_last.ckpt'.
| Missing keys: 0, Unexpected keys: 0
| loaded 'dit' from './checkpoints\diffusion_transformer/model_only_last.ckpt'.
| Missing keys: 0, Unexpected keys: 9
| loaded 'model' from './checkpoints\aligner_lm/model_only_last.ckpt'.
| Missing keys: 0, Unexpected keys: 0
Traceback (most recent call last):
  File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 267, in <module>
    infer_ins = MegaTTS3DiTInfer()
  File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 90, in __init__
    self.build_model(self.device)
  File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 144, in build_model
    self.g2p_model = AutoModelForCausalLM.from_pretrained(self.g2p_exp_name).eval().to(device)
  File "E:\anaconda\envs\megatts_win\lib\site-packages\transformers\models\auto\auto_factory.py", line 564, in from_pretrained
    return model_class.from_pretrained(
  File "E:\anaconda\envs\megatts_win\lib\site-packages\transformers\modeling_utils.py", line 3792, in from_pretrained
    if metadata.get("format") == "pt":
AttributeError: 'NoneType' object has no attribute 'get'

[jiang-ziyue]

@nightt5879 可以试试这个

[nightt5879]

@jiang-ziyue 我尝试了一下 前面的numpy确实没警告了，但是它后面还是报错这个get不到。我正在尝试解决中，我第一次是配置好的了 现在给我整迷惑了，感谢你的回复 我会继续尝试的

(megatts_win) E:\robot\project\MegaTTS3>pip install numpy==1.26.0
Looking in indexes: https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple, https://pypi.ngc.nvidia.com
Collecting numpy==1.26.0
  Downloading https://mirrors.tuna.tsinghua.edu.cn/pypi/web/packages/cc/05/ef9fc04adda45d537619ea956bc33489f50a46badc949c4280d8309185ec/numpy-1.26.0-cp310-cp310-win_amd64.whl (15.8 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 15.8/15.8 MB 1.9 MB/s eta 0:00:00
Installing collected packages: numpy
Successfully installed numpy-1.26.0

(megatts_win) E:\robot\project\MegaTTS3>python tts/infer_cli.py --input_wav 'assets/chinese_prompt.wav' --input_text "你好世界，我是中文测试" --output_dir ./gen --p_w 2.0 --t_w 3.0
| loaded 'dur_model' from './checkpoints\duration_lm/model_only_last.ckpt'.
| Missing keys: 0, Unexpected keys: 0
| loaded 'dit' from './checkpoints\diffusion_transformer/model_only_last.ckpt'.
| Missing keys: 0, Unexpected keys: 9
| loaded 'model' from './checkpoints\aligner_lm/model_only_last.ckpt'.
| Missing keys: 0, Unexpected keys: 0
Traceback (most recent call last):
  File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 267, in <module>
    infer_ins = MegaTTS3DiTInfer()
  File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 90, in __init__
    self.build_model(self.device)
  File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 144, in build_model
    self.g2p_model = AutoModelForCausalLM.from_pretrained(self.g2p_exp_name).eval().to(device)
  File "E:\anaconda\envs\megatts_win\lib\site-packages\transformers\models\auto\auto_factory.py", line 564, in from_pretrained
    return model_class.from_pretrained(
  File "E:\anaconda\envs\megatts_win\lib\site-packages\transformers\modeling_utils.py", line 3792, in from_pretrained
    if metadata.get("format") == "pt":
AttributeError: 'NoneType' object has no attribute 'get'

[nightt5879]

@jiang-ziyue 我朋友用我的流程可以跑，我这两天整理一下

[jiang-ziyue]

@nightt5879 好的，期待❤️

[forwiat]

@nightt5879 你可以尝试用以下代码转换一下g2p的model.safetensors：

import safetensors
from safetensors.torch import save_file

tensors = dict()
safetensors_path = "model.safetensors"
with safetensors.safe_open(safetensors_path, framework="pt") as f:
    for key in f.keys():
        tensors[key] = f.get_tensor(key)

save_file(tensors, safetensors_path, metadata={'format': 'pt'})

[zhanglifan1022nice]

@jiang-ziyue 我尝试了一下 前面的numpy确实没警告了，但是它后面还是报错这个get不到。我正在尝试解决中，我第一次是配置好的了 现在给我整迷惑了，感谢你的回复 我会继续尝试的

(megatts_win) E:\robot\project\MegaTTS3>pip install numpy==1.26.0
Looking in indexes: https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple, https://pypi.ngc.nvidia.com
Collecting numpy==1.26.0
Downloading https://mirrors.tuna.tsinghua.edu.cn/pypi/web/packages/cc/05/ef9fc04adda45d537619ea956bc33489f50a46badc949c4280d8309185ec/numpy-1.26.0-cp310-cp310-win_amd64.whl (15.8 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 15.8/15.8 MB 1.9 MB/s eta 0:00:00
Installing collected packages: numpy
Successfully installed numpy-1.26.0

(megatts_win) E:\robot\project\MegaTTS3>python tts/infer_cli.py --input_wav 'assets/chinese_prompt.wav' --input_text "你好世界，我是中文测试" --output_dir ./gen --p_w 2.0 --t_w 3.0
| loaded 'dur_model' from './checkpoints\duration_lm/model_only_last.ckpt'.
| Missing keys: 0, Unexpected keys: 0
| loaded 'dit' from './checkpoints\diffusion_transformer/model_only_last.ckpt'.
| Missing keys: 0, Unexpected keys: 9
| loaded 'model' from './checkpoints\aligner_lm/model_only_last.ckpt'.
| Missing keys: 0, Unexpected keys: 0
Traceback (most recent call last):
File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 267, in
infer_ins = MegaTTS3DiTInfer()
File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 90, in init
self.build_model(self.device)
File "E:\robot\project\MegaTTS3\tts\infer_cli.py", line 144, in build_model
self.g2p_model = AutoModelForCausalLM.from_pretrained(self.g2p_exp_name).eval().to(device)
File "E:\anaconda\envs\megatts_win\lib\site-packages\transformers\models\auto\auto_factory.py", line 564, in from_pretrained
return model_class.from_pretrained(
File "E:\anaconda\envs\megatts_win\lib\site-packages\transformers\modeling_utils.py", line 3792, in from_pretrained
if metadata.get("format") == "pt":
AttributeError: 'NoneType' object has no attribute 'get'

要更新transformer了。我也遇到同样的问题，通过以下解决了
x-transformers==1.44.4
transformers==4.49.0