Windows + AMD GPU, Finally Can Train Models with PyTorch

Preface

In the previous article, I started tinkering with compiling a ROCm build of PyTorch on Windows 11. The attempt failed, but I picked up some experience along the way.

And sure enough, the very next day the build succeeded.

Build Artifacts

So good!

-a---            2026/1/5    15:58      151148776 torch-2.9.1+rocm7.11.0a20260104-cp312-cp312-win_amd64.whl
-a---            2026/1/5    15:59         514989 torchaudio-2.9.0+rocm7.11.0a20260104-cp312-cp312-win_amd64.whl
-a---            2026/1/5    16:01        1247290 torchvision-0.24.0+rocm7.11.0a20260104-cp312-cp312-win_amd64.whl

Windows + AMD + ROCm + PyTorch,

Do you even get how valuable it is to have these words together?!! 😄

Key Points

To sum it up, there were just two key points to this successful build.

Use the Official Version of ROCm

Previously, because everything I found online said AMD still hadn't released ROCm for gfx103x, I installed a third-party prebuilt version. Maybe it was too old or whatever, but compiling PyTorch kept failing.

This time I installed with this command:

pip install --index-url https://rocm.nightlies.amd.com/v2-staging/gfx103X-dgpu/ --pre rocm[libraries,devel]

You can try looking for the corresponding ROCm packages at these locations:

• https://rocm.nightlies.amd.com/v2-staging/
• https://rocm.nightlies.amd.com/v2/
• https://rocm.devreleases.amd.com/v2
• https://rocm.devreleases.amd.com/v2-staging

Don't put blind faith in third-party packages. Use the official ones if you can; if not, compile ROCm yourself. Luckily my card has an official ROCm package.

Use an Older Version of PyTorch

I mentioned this in the previous article too: the latest version might not work. I saw someone online who successfully compiled version 2.10.

I actually had a vague sense of this while struggling last time: the newer the PyTorch version, the more likely you are to hit pitfalls on the Windows + ROCm path.

People online have already confirmed:

• 2.10: someone succeeded
• 2.11: various uncertainties exist

This time I went straight for the more conservative option: PyTorch 2.9.1

Turns out this was the right call.

As for whether 2.11 works?

I'll mess with that later if I have the energy. For now, I'll just enjoy the results.

Installing the Wheels

Now that it's compiled, I can install and try it out.

Install these three files in order (note: torch must be installed first)

pip install .\torch-2.9.1+rocm7.11.0a20260104-cp312-cp312-win_amd64.whl
pip install .\torchvision-0.24.0+rocm7.11.0a20260104-cp312-cp312-win_amd64.whl
pip install .\torchaudio-2.9.0+rocm7.11.0a20260104-cp312-cp312-win_amd64.whl

Windows + AMD GPU: the training pipeline is now open.

What's Next?

Now that GPU acceleration works, there's a lot I can do.

Get deep learning running, start playing with AI image generation~

Next I want to try:

• Some simple deep learning model training tasks
• Getting back into AI image generation
• Especially trying Z-Image : https://huggingface.co/Comfy-Org/z_image_turbo

There's still a lot worth documenting and sharing in this series. Stay tuned.