Support AMD GPUs on Windows

[34j]

Is your feature request related to a problem? Please describe. AMD GPUs not supported on Windows

Describe the solution you'd like AMD GPUs not supported on Windows

Additional context

[pierluigizagaria]

I'm trying to get this version working. I've installed the CPU version of torch because there is not installation for ROCm on Win. After installing with pip install -U git+https://github.com/34j/so-vits-svc-fork.git@feat/openml this error occurs after using svc train

Traceback (most recent call last):
  File "C:\Python310\lib\runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "C:\Python310\lib\runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "D:\Files\Code\python\so-vits-svc-fork\venv\Scripts\svc.exe\__main__.py", line 7, in <module>
  File "D:\Files\Code\python\so-vits-svc-fork\venv\lib\site-packages\click\core.py", line 1130, in __call__
    return self.main(*args, **kwargs)
  File "D:\Files\Code\python\so-vits-svc-fork\venv\lib\site-packages\click\core.py", line 1055, in main
    rv = self.invoke(ctx)
  File "D:\Files\Code\python\so-vits-svc-fork\venv\lib\site-packages\click\core.py", line 1657, in invoke
    return _process_result(sub_ctx.command.invoke(sub_ctx))
  File "D:\Files\Code\python\so-vits-svc-fork\venv\lib\site-packages\click\core.py", line 1404, in invoke
    return ctx.invoke(self.callback, **ctx.params)
  File "D:\Files\Code\python\so-vits-svc-fork\venv\lib\site-packages\click\core.py", line 760, in invoke
    return __callback(*args, **kwargs)
  File "D:\Files\Code\python\so-vits-svc-fork\venv\lib\site-packages\so_vits_svc_fork\__main__.py", line 130, in train
    train(config_path=config_path, model_path=model_path)
  File "D:\Files\Code\python\so-vits-svc-fork\venv\lib\site-packages\so_vits_svc_fork\train.py", line 41, in train
    raise RuntimeError("CUDA is not available.")
RuntimeError: CUDA is not available.

After commenting the two lines in train.py

    #if not torch.cuda.is_available():
        #raise RuntimeError("CUDA is not available.")

This is the output of the command. The training does not start at all.

(venv) PS D:\Files\Code\python\so-vits-svc-fork> svc train
[17:19:29] INFO     [17:19:29] Version: 1.3.3                                                                                                                                                   __main__.py:49
Downloading D_0.pth: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 178M/178M [00:04<00:00, 41.2MiB/s]
Downloading G_0.pth: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 172M/172M [00:03<00:00, 47.8MiB/s]
(venv) PS D:\Files\Code\python\so-vits-svc-fork> 

[34j]

Could you remove that part, pip install torch-openml and try again?

[pierluigizagaria]

Nothing found, I tried with pip install openml-pytorch but nothing changed

[34j]

Understood. I failed.

[pierluigizagaria]

Sure I did it, in fact no CUDA error is printed, but the command does nothing, it ends straight away

[34j]

https://github.com/34j/so-vits-svc-fork/blob/d7ab45010e73e63b765e9bd24c2635c2d18239be/src/so_vits_svc_fork/utils.py#L31 Can you check if this part is working? print(devices)

[34j]

Did svc pre-hubert work correctly?

[pierluigizagaria]

Yes hubert worked but no GPU was used

[pierluigizagaria]

But in this branch realtime inference works on GPU

edit. Not really

[34j]

@allcontributors add pierluigizagaria userTesting

[allcontributors[bot]]

@34j

I've put up a pull request to add @pierluigizagaria! :tada:

[34j]

It seems difficult to support, so I give up.

[pierluigizagaria]

Can this project work with 3.11? I'm trying to install torch-mlir that should make torch compatible my AMD GPU on Windows

I've already tried using torch-directml but got the error mentioned here https://github.com/microsoft/DirectML/issues/400

[34j]

> pipdeptree --reverse --packages llvmlite
Warning!!! Possibly conflicting dependencies found:
* poetry==1.4.2
 - platformdirs [required: >=2.5.2,<3.0.0, installed: 3.2.0]
------------------------------------------------------------------------
Warning!! Cyclic dependencies found:
* poetry-plugin-export => poetry => poetry-plugin-export
* poetry => poetry-plugin-export => poetry
------------------------------------------------------------------------
llvmlite==0.39.1
  - numba==0.56.4 [requires: llvmlite>=0.39.0dev0,<0.40]
    - librosa==0.9.1 [requires: numba>=0.45.1]
      - so-vits-svc-fork==3.0.4 [requires: librosa]
      - torchcrepe==0.0.18 [requires: librosa==0.9.1]
        - so-vits-svc-fork==3.0.4 [requires: torchcrepe>=0.0.17]
    - resampy==0.4.2 [requires: numba>=0.53]
      - librosa==0.9.1 [requires: resampy>=0.2.2]
        - so-vits-svc-fork==3.0.4 [requires: librosa]
        - torchcrepe==0.0.18 [requires: librosa==0.9.1]
          - so-vits-svc-fork==3.0.4 [requires: torchcrepe>=0.0.17]
      - scikit-maad==1.3.12 [requires: resampy>=0.2]
        - so-vits-svc-fork==3.0.4 [requires: scikit-maad]
      - torchcrepe==0.0.18 [requires: resampy]
        - so-vits-svc-fork==3.0.4 [requires: torchcrepe>=0.0.17]

3.10 is not supported for the above reasons, but won't it work with 3.11?

[pierluigizagaria]

I got an error while trying to install on 3.11

[34j]

Sorry, my typo, I was trying to ask if torch-mlir would work with 3.10.

[pierluigizagaria]

They don't provide compiled Windows versions on 3.10

[34j]

Since both inference and training rely on librosa as of now, 3.11 support is not possible.

[34j]

Installing the rc version of numba may allow it to be used with Python 3.11, but may cause other problems (I haven't tried it) (https://github.com/numba/numba/issues/8841)

[robonxt]

Would it be possible to run this using directml? (although I've only gotten directml to work on python 3.10.6, haven't tried it on newer versions)

[mikeyang01]

microsoft/DirectML#400

any update on this?