Configure SLURM resources and cluster environment

[esaran1]

Why

hpc/slurm_array_template.sh has generic laptop/HPC defaults (4h, 8 CPUs, 32G, GPU directive commented). Those are not a verified allocation for any named cluster.

What to do

On the actual cluster:

1. Fill in real #SBATCH time, CPUs, memory, partition, account, and GPU if needed.
2. Set ASLA_MODULES and ASLA_CONDA_ENV (or equivalent) so jobs find the same Python as local tests.
3. Confirm python scripts/hpc_preflight.py still passes on the login node.
4. Submit a dry-run array first:

sbatch --export=ALL,ASLA_SITE_COMMAND_TEMPLATE=hpc/site_command.template,ASLA_REAL_TRAIN_EVAL=/path/to/your_train_and_eval.py hpc/slurm_array_template.sh

5. Inspect logs/ and results/hpc/*/command.txt.
6. Only then submit with ASLA_EXECUTE=1.

The array must remain 1-N where N is the current manifest length (currently 45).

Acceptance

• One dry-run array job renders every row without training
• Environment modules/conda are documented in hpc/ADAPTATION_CHECKLIST.md
• Resource requests are enough for the defined recipes, not copied blindly
• Execute mode is still opt-in via ASLA_EXECUTE=1

Files

• hpc/slurm_array_template.sh
• hpc/ADAPTATION_CHECKLIST.md
• data/HPC_CHECKLIST.md

Do not fabricate

• Enable --execute in the template by default
• Change array bounds without regenerating the manifest