slurm-info-summary（slurm 信息摘要）

SLURM Info Summary

Collect SLURM cluster specs and save a polished, human-readable reference document.

Steps

1. Check for existing doc: Look for ~/.claude/skills/slurm-info-summary/references/slurm-cluster-summary.md.
2. If the doc already exists:

- Tell the user: "SLURM cluster summary already exists at ~/.claude/skills/slurm-info-summary/references/slurm-cluster-summary.md." - Read the file and display its content. - Do NOT re-run the script. Stop here.

1. If the doc does NOT exist:

- Run ~/.claude/skills/slurm-info-summary/scripts/gather-slurm-info.sh and capture stdout. - Parse the raw output (structured with === SECTION === markers) and produce a polished markdown summary following the template below. - Write the summary to ~/.claude/skills/slurm-info-summary/references/slurm-cluster-summary.md. - Display the summary to the user. - Tell the user the file path where it was saved.

Output Template

Use the raw data to produce a summary that matches this structure and style exactly. Convert raw memory values from MB to human-readable GB/TB. Derive node types by grouping nodes with the same prefix (e.g. ravc, ravg, ravh, ravl).

# <ClusterName> Cluster Overview

> Auto-generated on <UTC timestamp> by `/slurm-info-summary`

All compute nodes use **<CPU model>** processors with **<sockets> sockets, <cores>/socket, <threads> threads/core = <total logical CPUs> CPUs** per node.

---

## Partitions

| Partition | Nodes | Node Type | Memory/Node | GPUs/Node | Max Walltime | Max Nodes/Job | Oversubscribe |
|-----------|-------|-----------|-------------|-----------|--------------|---------------|---------------|
| ... | ... | ... | ... | ... | ... | ... | ... |

---

## Node Types

| Prefix | Count | Memory | GPUs | Notes |
|--------|-------|--------|------|-------|
| ... | ... | ... | ... | ... |

---

## Key Partition Differences

- **`<partition_a>` vs `<partition_b>`**: <explain the difference concisely>
- ...

---

## QOS Limits (notable only)

| QOS | Max Nodes/Job | Max Running Jobs | Max Submit Jobs | Max Walltime |
|-----|---------------|------------------|-----------------|---------------|
| ... | ... | ... | ... | ... |

Only include QOS entries that have at least one non-empty limit.

---

## Usage Examples

Provide 5-7 ready-to-use `sbatch`/`srun` examples covering:
- Interactive session
- Single-node CPU job (small partition)
- Multi-node CPU job (general partition)
- Single-GPU shared job (gpu1 partition)
- Multi-node GPU exclusive job (gpu partition)
- Quick GPU dev/test (gpudev partition)
- High-memory node request (if available)

## Key Tips

- Bullet list of practical tips: billing weights, constraint flags, useful commands (`squeue`, `scancel`), etc.

Important

• Do NOT output the raw script data to the user. Only output the polished summary.
• Keep the summary concise but complete.
• The "Node Availability" section from the script is point-in-time data — do NOT include it in the saved summary (it would be stale).
• Physical cores vs logical CPUs: Nodes with hyperthreading have more logical CPUs than physical cores (e.g., 72 physical cores = 144 logical CPUs with 2 threads/core). SLURM's --cpus-per-task counts physical cores. When describing per-GPU resource limits for shared partitions, always state the value in physical cores and note the logical CPU count parenthetically. For example: "18 physical cores (36 logical CPUs) and 125 GB memory per GPU".