Stable Diffusion

Affirm.

ROCm is basically AMD's answer to CUDA. Just (as usual) more open, less polished, and harder to use. Using something called HIP, CUDA application can be translated to work with ROCm instead (and therefore run on AMD cards without a complete rewrite of the app).

AFAIK they started working on it 6 or 7 years ago as the replacement for OpenCL. Not sure why exactly, but OpenCL apparently wasn't getting enough traction (and I think Blender even recently dropped OpenCL support).

After all the time, the HW support is still spotty (mostly only supporting the Radeon Pro cards, and still having no proper support for RDNA3 I think), and the SW support focuses mainly on Linux (and only three blessed distros, Ubuntu, RHEL and SuSe get official packages, so it can be pain to install anywhere else due to missing or conflicting dependencies).

So ROCm basically does work, and keeps getting better, but nVidia clearly has a larger SW dev team that makes the CUDA experience much more polished and painless.

Rocm is the Radeon Open Compute ... Something begining with m. It's the libraries that AMD write to divert workloads to your GPU.

I have a 7600 and I'm finding things a bit bleeding edge (on Linux). I highly recommend making sure you're on the latest 7.6.1 of Rocm and make sure the version of pytorch you're using is a recent nightly that uses 7.6. earlier versions didn't support the 7xxx series.

If you're on windows, I think thinks can work differently.

Oh boy you're in for a nice learning time.