AMD (ROCm) support through pytorch

[olinesn]

Hello,

Now that sleap 1.5 uses pytorch, I think it's time to start testing non-nVidia (CUDA) hardware for both training and inference.

It appears that pytorch supports AMD (ROCm), but only on linux, if I'm interpreting the command tool correctly (https://pytorch.org/).

I purchased a GMKtec EVO-X2, which has an integrated AMD gpu. I think this would be an ideal system to test on, because AMD is investing development support into this specific processor (AMD Ryzen AI Max+ 395), and pytorch is reportedly supported and working on the iGPU using ROCm by other users, and from official sources.

As a first pass test, I've installed Ubunto on my EVO X2, and then run a modified version of the sleap install command, based on the pytorch homepage commmand tool.

uv tool install --python 3.13 "sleap[nn]" --index-url https://download.pytorch.org/whl/rocm6.4 --index https://pypi.org/

This executes correctly, but when I then open a sleap project in the GUI and initialize training, the command line lists that there are no GPUs found, and the training runs on the CPU (loudly!).

Could I please ask for support for next steps? If we figure it out, I think we could have a nice set of instructions recommending specific computers (that are NOT DELL).

Thanks!