gpu-preflight

name	gpu-preflight
description	Check idle GPU availability across both nodes (gpu01 and gpu02) and recommend candidate devices before launching jobs. Use when you need to choose a node/GPU for training or inference, or to report GPU availability with a concrete recommendation.

GPU Preflight

Quick start

python3 ~/.codex/skills/gpu-preflight/scripts/gpu_preflight.py

What this skill does

• Queries GPU status on gpu01 and gpu02 (8 GPUs each).
• Summarizes per-GPU utilization and memory usage.
• Produces a short list of candidate GPUs and a single recommended node:gpu_index.

Rules

• Always check both nodes (gpu01, gpu02) before making any GPU scheduling recommendation.
• Prefer ssh in batch mode and fail fast if a node is unreachable (do not hang waiting for a password prompt).

CLI usage

python3 ~/.codex/skills/gpu-preflight/scripts/gpu_preflight.py --help

Common options:

• --nodes gpu01,gpu02 (override node list)
• --max-util 5 (candidate threshold, percent)
• --max-mem-used-mib 1000 (candidate threshold)
• --top-k 5 (how many candidates to print)
• --json (machine-readable output)

Output contract (for agent reporting)

• When reporting availability, include both nodes and the recommended GPU candidates.
• If any node cannot be queried, state that explicitly and do not assume it is idle.