[JRPan]

Retires accel-sim/Dockerfile. Its image branches move here, one file per image, and are built by a new workflow.

What moves

• Each image branch becomes dockerfiles/<tag>.Dockerfile, built as ghcr.io/accel-sim/accel-sim-framework:<tag>.
• Ubuntu-24.04-cuda-13.2-minimal comes from Ubuntu 24.04 CUDA 13.2.1 minimal Dockerfile#12. The branch of that name held a 12.8 Dockerfile (FROM ...cuda:12.8.0).
• Ubuntu-24.04-cuda-12.8-minimal gains the AWS CLI and boto3, so CI's image can also run AWS Batch jobs.
• Left out, both never published: Ubuntu-18.04-cuda-11.0 and Ubuntu-22.04-cuda-11.7, from 2025.
• Source labels now point at this repo.

When images build

.github/workflows/docker-images.yml builds only the Dockerfiles a change touches. The rules are in .github/scripts/docker-plan.sh, which logs a build/skip reason for each image to the job summary:

• Pull request: builds the changed images, without pushing.
  • A modified file always builds.
  • An added file builds only if its tag is not on ghcr yet, so this import rebuilds nothing that already exists.
• Merge to dev: builds again and pushes. pc02 has the PR build's layers cached, so the push is the only real work.
• workflow_dispatch: builds the tags you name, with an optional no_cache.
• The jobs run only in accel-sim/accel-sim-framework.

Builds run on tgrogers-raid, which has no docker daemon. The runner account's docker CLI and buildx drive tgrogers-pc02's daemon over DOCKER_HOST=ssh://tgrogers-pc02. The same setup published the current 12.8-minimal from accel-sim/Dockerfile (run 41171899169).

Keeping the apps current

The workflow passes gpu-app-collection dev's HEAD as the GPU_APP_COLLECTION_SHA build arg, so the cached clone layer rebuilds only when dev moves. Without it, tgrogers-pc02 would keep reusing the first clone it made.

Expected on this PR

Two full builds, Ubuntu-24.04-cuda-13.1.1-minimal and Ubuntu-24.04-cuda-13.2-minimal, which are not on ghcr yet. The other 7 images are skipped. Both need accel-sim/gpu-app-collection#91, now merged.

Before merge

The ghcr package accel-sim-framework is still linked to accel-sim/Dockerfile. In its package settings:

1. Unlink accel-sim/Dockerfile and connect this repo.
2. Give this repo Write under Manage Actions access.

Otherwise the post-merge push is denied.

Testing

• docker-plan.sh was checked against real history:
  • this import builds only 13.1.1 and 13.2
  • modify / delete / add
  • no change
  • first push of a branch
  • dispatch, including a bad tag
• All the Dockerfiles pass docker buildx build --call=check with dockerfiles/ as the context.

🤖 Generated with Claude Code