在执行ONNX导出时无法处理rmvpe.py中的Encoder BN

Hi:
我正在尝试将vc_single()函数封装成一个torch.nn.Module模型类，现在已经做到了，实现的notebook在这里。
可问题是，当我尝试把这个Pipeline使用torch.onnx.export导出为ONNX模型时，它抛出了如下的错误：

RuntimeError: Cannot insert a Tensor that requires grad as a constant. Consider making it a parameter or input, or detaching the gradient
Tensor:
0.6670
[ torch.cuda.HalfTensor{1} ]

通过观察StackTrace我发现报错的代码在此处：

File ~/Retrieval-based-Voice-Conversion-WebUI-main/infer/lib/rmvpe.py:244, in Encoder.forward(self, x)
  [242](https://vscode-remote+ssh-002dremote-002bconnect-002enmb1-002eseetacloud-002ecom.vscode-resource.vscode-cdn.net/root/Retrieval-based-Voice-Conversion-WebUI-main/~/Retrieval-based-Voice-Conversion-WebUI-main/infer/lib/rmvpe.py:242) def forward(self, x: torch.Tensor):
  [243](https://vscode-remote+ssh-002dremote-002bconnect-002enmb1-002eseetacloud-002ecom.vscode-resource.vscode-cdn.net/root/Retrieval-based-Voice-Conversion-WebUI-main/~/Retrieval-based-Voice-Conversion-WebUI-main/infer/lib/rmvpe.py:243)     concat_tensors: List[torch.Tensor] = []
--> [244](https://vscode-remote+ssh-002dremote-002bconnect-002enmb1-002eseetacloud-002ecom.vscode-resource.vscode-cdn.net/root/Retrieval-based-Voice-Conversion-WebUI-main/~/Retrieval-based-Voice-Conversion-WebUI-main/infer/lib/rmvpe.py:244)     x = self.bn(x)
  [245](https://vscode-remote+ssh-002dremote-002bconnect-002enmb1-002eseetacloud-002ecom.vscode-resource.vscode-cdn.net/root/Retrieval-based-Voice-Conversion-WebUI-main/~/Retrieval-based-Voice-Conversion-WebUI-main/infer/lib/rmvpe.py:245)     for i, layer in enumerate(self.layers):
  [246](https://vscode-remote+ssh-002dremote-002bconnect-002enmb1-002eseetacloud-002ecom.vscode-resource.vscode-cdn.net/root/Retrieval-based-Voice-Conversion-WebUI-main/~/Retrieval-based-Voice-Conversion-WebUI-main/infer/lib/rmvpe.py:246)         t, x = layer(x)

请问有什么办法可以解决此问题吗？因为从Encoder类的定义上看似乎没看出哪里有问题

Start with the POC_Torch.ipynb reproduction and inspect infer/lib/rmvpe.py, especially Encoder.forward() at the self.bn(x) call and the surrounding Encoder initialization. Run the torch.onnx.export path from the notebook and trace how the BN state is registered; done means the pipeline exports without the reported requires-grad constant error.