Sage → Apptainer → Slurm pipeline

Commit Sage code, GitHub Actions builds a self-contained Apptainer image, and the cluster pulls and runs it as a Slurm job.

Repository layout

.github/workflows/build-sage-container.yml   CI workflow
container/sage.def                           Apptainer definition
code/main.sage                               Your Sage entry point (edit me)
slurm/run_sage.sbatch                        Submission script for the cluster

How it works

1. A push to main that touches code/ or container/ triggers the workflow.
2. The runner installs Apptainer, builds sage-job.sif from container/sage.def (which bakes code/ into the image at /opt/sage-code), and runs apptainer test as a smoke test.
3. The SIF is pushed to GitHub Container Registry as an ORAS artifact, tagged latest and with the commit SHA. It's also attached as a workflow artifact as a fallback.
4. On the cluster, sbatch slurm/run_sage.sbatch pulls the image and runs it. The container's runscript executes sage /opt/sage-code/main.sage, forwarding any arguments you pass to sbatch ... script args.

One-time setup

1. Edit REPO in slurm/run_sage.sbatch to your ghcr.io/<user>/<repo>/sage-job path.

2. After the first push, go to the package on GitHub (Profile → Packages → sage-job) and either make it public (simplest for the cluster) or, if it must stay private, authenticate on the cluster once with a classic PAT that has read:packages:

   apptainer registry login --username <gh-user> oras://ghcr.io
   

3. Make sure your cluster's login/compute nodes can reach ghcr.io. If they can't, download the workflow artifact instead and scp the SIF over.

Running on the cluster

sbatch slurm/run_sage.sbatch                # runs main.sage
sbatch slurm/run_sage.sbatch --n 1000      # args are forwarded to main.sage
SAGE_IMAGE_TAG=<commit-sha> sbatch slurm/run_sage.sbatch   # pin a specific build

Outputs written to /output inside the container land in results-<jobid>/ next to where you submitted.

Notes and knobs

• Image size: the image installs core Sage from conda-forge on a micromamba base — typically ~2 GB instead of the ~6+ GB official Docker image. Need extra Python packages? Add them to the micromamba install line in container/sage.def (e.g. numpy scipy pandas).
• Reproducibility: pin the Sage version in sage.def (pinned to sage=10.4 in the conda install) and submit jobs with the commit-SHA tag rather than latest.
• Different entry point: change the %runscript in sage.def, or override at runtime with apptainer exec sage-job.sif sage /opt/sage-code/whatever.sage.
• MPI / GPU: add --nv to the apptainer run line for CUDA, or restructure with srun apptainer exec ... for multi-task MPI jobs.