I built this and the other rocm packages from AUR (some removed now) along with pytorch build-flagged for rocm. It did eventually work, but It was a real pain in the ass and cost a whole lot of time. After that though, something broke (who knows) and I ended up rebuilding everything again just to get it back to a working state. I'm currently too afraid to update to these community packages, partially because I don't know if the build flags I used are the duct tape keeping my builds functional (I have an rx590-8gb which is maybe at the edge of supported gpus). Instead, I just added every rocm package to my IgnorePkgs, which isn't a great sign.

Unfortunately, ROCm is broken for gfx803 (and possibly others?). I don't think they perform regression testing with their older cards.

I think this is due to custom assembly kernels that are now outdated, I recall seeing somewhere that removing the optimized kernels and allowing to fall back to the generic implementation allowed it to work.

ROCm is in such a sad state. If I can't make it work at home, why would I ever want to investigate it at work?