Slurm Workflows
Trigger Boundary
Use this skill for Slurm or HPC batch work that needs resource planning, script generation, submission review, job arrays, queue/status interpretation, log triage, failed-job diagnosis, or safe resubmission.
Do not use it for generic shell scripting, non-Slurm schedulers, or site-specific partition/account decisions unless a site profile or local override provides those constraints.
Required Site Context
Before giving final Slurm commands, read the generated site profile reference when present:
references/_generated/site-profile.md
If no site profile is installed, ask for or detect the minimum non-secret site facts: partitions, account/QOS policy, GPU availability, walltime limits, scratch/log conventions, module setup, and whether race execution is permitted.
Workflow
- Identify workload shape: CPU/GPU, memory, walltime, array size, input/output paths, scratch needs, dependency graph, and restartability.
- Apply site constraints from the profile. Do not invent partition, account, QOS, or hostname values.
- Produce a script with explicit logs, working directory, environment setup hook, resource requests, and failure-visible shell options.
- Prefer dry-run or header review before submission when the user is deciding resources.
- Monitor with current scheduler status and logs; distinguish queued, running, failed, cancelled, timed out, out-of-memory, dependency-held, and completed states.
- On failure, classify the cause and propose the smallest safe retry.
Race Policy
Race execution means submitting multiple alternative jobs and cancelling losers after a winner is verified. Only use it when the site profile explicitly allows race execution and the user approves the resource cost. Always record cancellation criteria.
Outputs
Return the job script or commands, expected log paths, monitoring commands, stop conditions, and final verification checks. Submission is not completion; completion requires current job state and final output validation.
1---2name: slurm-workflows3description: Plan, submit, monitor, diagnose, and safely iterate Slurm jobs with generic resource estimation, job arrays, log/scratch layout, queue inspection, failure classification, and optional race execution policy supplied by a site profile.4---5# Slurm Workflows67## Trigger Boundary89Use this skill for Slurm or HPC batch work that needs resource planning, script generation, submission review, job arrays, queue/status interpretation, log triage, failed-job diagnosis, or safe resubmission.1011Do not use it for generic shell scripting, non-Slurm schedulers, or site-specific partition/account decisions unless a site profile or local override provides those constraints.1213## Required Site Context1415Before giving final Slurm commands, read the generated site profile reference when present:1617`references/_generated/site-profile.md`1819If no site profile is installed, ask for or detect the minimum non-secret site facts: partitions, account/QOS policy, GPU availability, walltime limits, scratch/log conventions, module setup, and whether race execution is permitted.2021## Workflow22231. Identify workload shape: CPU/GPU, memory, walltime, array size, input/output paths, scratch needs, dependency graph, and restartability.242. Apply site constraints from the profile. Do not invent partition, account, QOS, or hostname values.253. Produce a script with explicit logs, working directory, environment setup hook, resource requests, and failure-visible shell options.264. Prefer dry-run or header review before submission when the user is deciding resources.275. Monitor with current scheduler status and logs; distinguish queued, running, failed, cancelled, timed out, out-of-memory, dependency-held, and completed states.286. On failure, classify the cause and propose the smallest safe retry.2930## Race Policy3132Race execution means submitting multiple alternative jobs and cancelling losers after a winner is verified. Only use it when the site profile explicitly allows race execution and the user approves the resource cost. Always record cancellation criteria.3334## Outputs3536Return the job script or commands, expected log paths, monitoring commands, stop conditions, and final verification checks. Submission is not completion; completion requires current job state and final output validation.
YuukiAS