invoke-ai
[Bugfix] Fix ROCm support in Docker container
libstdc++-10-dev is needed to fix a HIPRTC compilation issue when a prompt is submitted under certain parameters for the first time. I'm not sure if the results of this compilation can be cached to avoid compiling again later on.
A symbolic link for amdgpu.ids was also missing.
With this patch, creating and running the docker container as follows enables CUDA as expected. Only the proper kernel drivers on the host should be needed (tested using a Radeon RX 6900 XT):
CONTAINER_FLAVOR=rocm ./docker/build.sh
CONTAINER_FLAVOR=rocm ./docker/run.sh
@lstein are you able to test this? I do not have access to a system with an AMD GPU, and not sure if anyone else on the team does...
@Rubonnek thanks again for the contribution. we're refactoring the Docker setup quite heavily in #3587 . Would you be interested in reviewing / testing that PR and implementing your ROCm fixes based on those changes?
Thank you!
@ebr I'm quite busy lately and I won't be able to test/patch for a while, but I'll look into it once I find the time.
For now, I have no further intentions of maintaining this PR so I'm closing it.