CUDA library is a stub, had to remove /usr/local/cuda/lib64/stubs about stable-diffusion-docker HOT 7 CLOSED

20.10.17, on NixOS. Which honestly I thought was the cause of my problems; nix with python stuff has been tricky. But this one was actually docker related, possibly a bug in tensorflow's latest-gpu image?

from stable-diffusion-docker.

I will give it a try later tonight and get back to you.

from stable-diffusion-docker.

Built on tensorflow:2.9.2-gpu, could run without deleting that dir. However, went back to :latest, and could also run... will need to do a more rigorous test, could be that I've polluted my docker builds and whatnot. Will try to reproduce the issue and get back to you.

from stable-diffusion-docker.

Ok, thoroughly confused now.. getting pretty inconsistent behavior while I try various tags of tensorflow. Sometimes the /usr/local/cuda directory exists, in which case /usr/local/cuda/lib64/stubs needs to be deleted for CUDA to load, other times /usr/locall/cuda doesn't exist at all, in which case CUDA loads properly. Also strangely, rebuilds of the stable-diffusion-docker image seem to work without needing to delete the stubs directory, regardless of which tensorflow tag it's based on.

Altogether baffling, my docker-fu is clearly not strong enough. Given that no one else seems to be having this issue, probably safe to close.

from stable-diffusion-docker.

It might be worth adding this to the README. Which version of Docker are you using?

from stable-diffusion-docker.

If you try to build the image on an older version of tensorflow with gpu, do you still get the error? If so I might switch to using specific versions of the images.

from stable-diffusion-docker.

Closing the issue for now, it can always be reopened.

from stable-diffusion-docker.