I’m using Ubuntu(to follow what AMD has for ROCm) and building the entirety of (gfx803 patched) ROCm from source.
It works with some forks but not others.
I also tried to compile pytorch with its Vulkan backend, but ended throwing the towel as LDFLAGS are a mess to get right (I successfully compiled it, but that was only part of the build chain, and decided I had better things to spend time on). I wonder how that would perform; ncnn works pretty decently.
It works with an older fork that has img2img / txt2img separated from each other, but nothing else did.
When you built it, you did at least source the env.sh to pick up the build environment that it needed?
It’s possible I had to do something weird, but I think it supports Polaris OOB. rocminfo certainly thinks so.
I’m running it on Debian testing, with an equivs package to get it to install cleanly:
$ cat rocm-equivs
Package: amdgpu-driver-fixes
Provides: python,libstdc++-7-dev,libgcc-7-dev
Architecture: all
Description: Fixes the AMD GPU driver installation on Debian testing
I had to compile my own version of pytorch to get gfx803 support there.I‘ll see if I can recreate the steps & create a runbook.