How to launch distributed Megatron-LM training jobs on a SLURM cluster. Covers a minimal sbatch skeleton, environment-variable setup for torch.distributed.run, CUDA_DEVICE_MAX_CONNECTIONS rules across hardware and parallelism modes, container conventions, monitoring, and per-rank failure diagnosis.
Install
npx skills add NVIDIA/skills --skill mcore-run-on-slurmnpx skills add NVIDIA/skills --skill mcore-run-on-slurm -a claude-code -gnpx skills add NVIDIA/skills --skill mcore-run-on-slurm -a codex -ggemini skills install https://github.com/NVIDIA/skills.git --path skills/mcore-run-on-slurmgh skill install NVIDIA/skills mcore-run-on-slurmgit clone --depth 1 https://github.com/NVIDIA/skills /tmp/NVIDIA__skills
mkdir -p ~/.claude/skills ~/.agents/skills
cp -R /tmp/NVIDIA__skills/skills/mcore-run-on-slurm ~/.claude/skills/mcore-run-on-slurm # Claude Code
cp -R /tmp/NVIDIA__skills/skills/mcore-run-on-slurm ~/.agents/skills/mcore-run-on-slurm # Codex, Cursor, Gemini CLI, Copilot, OpenCode