[Feature]: torch releases should include torchvision and torchaudio

[stellaraccident]

Suggestion Description

I remember that we scripted this but can't remember why it isn't in CI.

Users are asking for it.

Operating System

No response

GPU

No response

ROCm Component

No response

[marbre]

We already do, see for example https://d2awnip2yjpvqn.cloudfront.net/v2/gfx94X-dcgpu/torchaudio/ and https://d2awnip2yjpvqn.cloudfront.net/v2/gfx94X-dcgpu/torchvision/.

[ScottTodd]

On Windows we will at least need pytorch/audio#3945 for torchaudio. We can cherry-pick that for now but will want a solid plan for upstream support eventually too.

[rjygraham]

As an individual new to AMD, I grossly underestimated the impact of AMD vs NVIDIA in the world of local AI. This is exasperated by the fact my AMD system of choice is gfx1151 Strix Halo. Having said that, I want to thank you all for the unrelenting efforts on TheRock project - you are inspiring.

I've got ComfyUI up and running and can use it to run some workflows but most fail. I'd like to be on the bleeding edge of testing new TheRock releases for gfx1151 but I can't without the Windows torchvideo/torchaudio wheels that go along with the torch wheel.

I was going to try building them locally using the nighly builds but I went through the docs at the link below but it appears building torchvideo/torchaudio wheels for Windows is still not ready?
https://github.com/ROCm/TheRock/blob/main/external-builds/pytorch/README.md

Any guidance is appreciated.

[ScottTodd]

Working on enabling torch audio and vision on Windows now.

On Windows we will at least need pytorch/audio#3945 for torchaudio. We can cherry-pick that for now but will want a solid plan for upstream support eventually too.

I'm not sure if that patch is still needed for torch audio. I needed a different patch though, allowing clang-cl as the compiler instead of just cl, given that PyTorch sets a whole bunch of clang-specific flags when ROCm is enabled: #1091.

As for torch vision, I'm getting errors like

In file included from D:\projects\TheRock\external-builds\pytorch\.venv\Lib\site-packages\torch\include\torch/csrc/Device.h:4:
D:\projects\TheRock\external-builds\pytorch\.venv\Lib\site-packages\torch\include\torch/csrc/python_headers.h(12,10): warning: #include resolved using non-portable Microsoft search rules as: D:\projects\TheRock\external-builds\pytorch\.venv\Lib\site-packages\torch\include\torch\csrc\api\include\torch\Python.h [-Wmicrosoft-include]
   12 | #include <Python.h>
      |          ^
D:\projects\TheRock\external-builds\pytorch\.venv\Lib\site-packages\torch\include\torch/csrc/python_headers.h(12,10): error: 'Python.h' file not found with <angled> include; use "quotes" instead
   12 | #include <Python.h>
      |          ^~~~~~~~~~
      |          "Python.h"
D:\projects\TheRock\external-builds\pytorch\.venv\Lib\site-packages\torch\include\torch/csrc/python_headers.h(12,10): warning: non-portable path to file '<python.h>'; specified path differs in case from file name on disk [-Wnonportable-include-path]
   12 | #include <Python.h>
      |          ^~~~~~~~~~
      |          <python.h>
D:\projects\TheRock\external-builds\pytorch\.venv\Lib\site-packages\torch\include\torch/csrc/python_headers.h(13,10): fatal error: 'frameobject.h' file not found
   13 | #include <frameobject.h>
      |          ^~~~~~~~~~~~~~~

audio had this code: https://github.com/pytorch/audio/blob/b6a3368a45aaafe05f1a6a9f10c68adc5e944d9e/tools/setup_helpers/extension.py#L168-L172, but vision uses CppExtension directly, instead of a CMake build: https://github.com/pytorch/vision/blob/main/setup.py . I'll see if @jammm or @scottt had patches on a branch somewhere to get torchaudio building on Windows with ROCm...

[ScottTodd]

Ah, pytorch audio on main has pytorch/audio@bccaa45 with that clang-cl compatibility fix, mixed with some other cmake/LoadHIP.cmake changes to untangle. Worth checking pytorch vision too, now that I'm rebased onto main instead of older release branches: #1039.