Skip to content

Conversation

@CISC
Copy link
Collaborator

@CISC CISC commented Dec 28, 2025

Closes #18383

In addition to adding a new CUDA 13.1 image (with cuda13 tag) this also adds the cuda12 tag, the old cuda tag still points to CUDA 12.4 image, but can be changed in the future.

I tried to reuse the cuda.Dockerfile for both (hence the new build-args logic), but it ultimately failed due to --break-system-packages incompatibilities between Ubuntu 22.04 and 24.04 and the lack of a 24.04 image for CUDA 12.4.

Successful run for 12.4:
https://github.com/CISC/llama.cpp/actions/runs/20558563596/job/59046091086

Successful run for 13.1:
https://github.com/CISC/llama.cpp/actions/runs/20557901518/job/59044605518

@JohannesGaessler
Copy link
Collaborator

FYI you can circumvent --break-system-packages if you delete the EXTERNALLY_MANAGED file in a Python environment, that is what causes pip to throw a tantrum.

@CISC
Copy link
Collaborator Author

CISC commented Dec 28, 2025

FYI you can circumvent --break-system-packages if you delete the EXTERNALLY_MANAGED file in a Python environment, that is what causes pip to throw a tantrum.

Nice to know, though at some point we will have to phase out 12.4 so until then we can live with two Dockerfiles I think.

@am17an
Copy link
Collaborator

am17an commented Dec 29, 2025

Note that CUDA >=13 drops support for Pascal, so it may not be so easy to phase out

@CISC
Copy link
Collaborator Author

CISC commented Dec 29, 2025

Note that CUDA >=13 drops support for Pascal, so it may not be so easy to phase out

Oh sure, hopefully in favor of a newer 12.x, we will keep backward compatibility for as long as possible, I just don't want to change 12.4 until we have to, just in case.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

devops improvements to build systems and github actions

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Feature Request: Docker images for CUDA 12.8

4 participants