PyTorch usability on AMD hardware?

I would very much like to get an AMD GPU as my upcoming upgrade, but PyTorch support is crucial and I cannot find any report of successful application.

Can anyone provide feedback on PyTorch + AMD ROCm/HIP usability? Preferably on Linux.

My specific application is image denoising (UNet / cGAN) and I run Arch Linux but that’s mostly irrelevant.

As of today, not easy to build, nor as effective as tensorflow benchmarks suggest. I would not recommend it. I’m keeping an eye open for next months upcoming ameliorations (maybe ?)

New (I believe !) instructions here :

1 Like

As people link this thread: Building PyTorch for ROCm is quite easy: also AMD publishes docker images.

Best regards