ROCM Cuda not available
ironstarfish opened this issue · comments
ironstarfish commented
Describe the bug
Trying to run pytorch with CUDA. Can only run on CPU.
AMD GPU detected using rocminfo.
Tried to install rocm components on top, still get the issue.
What did you expect to happen?
CUDA capabilities enabled.
Output of rpm-ostree status
State: idle
Deployments:
● ostree-image-signed:docker://ghcr.io/ublue-os/bazzite-gnome:stable
Digest: sha256:1a7a9e5229cfe2ba234ba78292c411b3d53755f39dfe1fa77239e8367c745f2a
Version: 40.20240427.0 (2024-04-27T22:34:27Z)
LayeredPackages: bpytop htop rclone rocm-clinfo rocm-hip rocm-opencl
LocalPackages: lact-libadwaita-0.5.4-0.x86_64
ostree-image-signed:docker://ghcr.io/ublue-os/bazzite-gnome:stable
Digest: sha256:1a7a9e5229cfe2ba234ba78292c411b3d53755f39dfe1fa77239e8367c745f2a
Version: 40.20240427.0 (2024-04-27T22:34:27Z)
LayeredPackages: bpytop htop rclone
LocalPackages: lact-libadwaita-0.5.4-0.x86_64
Hardware
X470 AORUS GAMING 7 WIFI
Ryzen 5800X
Radeon 6900XT
Extra information or context
No response
Kyle Gospodnetich commented
ROCM is definitely working here on the same GPU. I would try unlayering all of that with rpm-ostree reset
, and then running pytorch via distrobox with ROCM's libraries installed in the box.