Partition

Partition

Use this skill when the user asks to choose, rank, compare, or change Slurm partitions for job scripts on the Berkeley cluster.

Core Workflow

1. Query current Slurm data before deciding. Partition availability changes.
2. Rank partitions using current availability, physical max memory, CPU count, GPU presence, account/QOS association, and benchmark suitability.
3. For given .sh job files, read their #SBATCH --mem request and choose a partition whose physical max memory is strictly larger than the request.
4. Edit only these SBATCH fields unless the user explicitly asks otherwise: --partition, --account, --qos.
5. Do not change --mem. If no ranked partition can satisfy the existing memory request, stop and explain the blocker.
6. After edits, print one representative modified .sh file or at least its SBATCH header so the user can verify the change.

Helper Script

Prefer the bundled helper for ranking and edits:

python3 /global/home/users/yaoshen/.codex/skills/partition/scripts/partition_helper.py rank
python3 /global/home/users/yaoshen/.codex/skills/partition/scripts/partition_helper.py update path/to/job.sh
python3 /global/home/users/yaoshen/.codex/skills/partition/scripts/partition_helper.py update --apply path/to/job1.sh path/to/job2.sh

update defaults to dry-run. Use --apply only when the user asked for file edits. The helper prints the selected partition/account/QOS and an example header. It queries Slurm with sacctmgr, sinfo, and scontrol; if sandboxing blocks Slurm sockets, rerun with escalated permissions.

Common options:

--cpu-only              Prefer CPU partitions and penalize GPU partitions.
--allow-gpu             Permit GPU partitions in the ranking.
--include-interactive   Permit ood_inter.
--prefer PARTITION      Add a ranking bonus to one partition.
--partition PARTITION   Force a partition if it fits the existing --mem.
--show-lines N          Print N lines from the example script.

Local Policy

For routine CPU-only WB97M(OS) or Q-Chem benchmarks, prefer CPU partitions over GPU partitions. Current practical defaults are usually:

• cm1, account=lr_qchem, qos=condo_qchem
• lr8, account=lr_mhg2, qos=mhg2_lr8_normal
• mhg, account=mhg, qos=normal
• lr6, account=lr_mhg2, qos=lr_lowprio
• lr7, account=lr_mhg2, qos=condo_mhg_lr7 by default; use another lr7 QOS only when explicitly requested by the user

Use lr_bigmem only when the existing memory request is above normal CPU-node capacity. Avoid ood_inter for production benchmark batches unless the user explicitly asks for interactive work.

Manual Editing Rules

If editing without the helper, preserve all unrelated lines and comments. Modify or insert exactly these lines near the top SBATCH block:

#SBATCH --partition=<partition>
#SBATCH --account=<account>
#SBATCH --qos=<qos>

Validate after editing with:

rg -n '^#SBATCH --(partition|account|qos|mem)=' path/to/job.sh
sed -n '1,40p' path/to/job.sh