Launch and manage GPU-accelerated SLURM jobs on the SONIC cluster.
npx skills add https://github.com/aaaxn/skills --skill slurm-aaaxn
Please help me install this Agent Skill. Skill: slurm Source: https://github.com/aaaxn/skills/tree/main/slurm Command: npx skills add https://github.com/aaaxn/skills --skill slurm-aaaxn
Streamlines launching and managing GPU-accelerated jobs on the SONIC SLURM cluster, handling environment setup, partition selection, and dataset staging to prevent common mismatches and timeouts.
Submit your first SLURM job from phocus4 after exporting SLURM_CONF and selecting an idle partition.
High-intent search queries and answers about installing and using this skill.
To submit a SLURM job on SONIC from phocus4, you must export SLURM_CONF before executing sbatch, enforce an idle partition, and stage datasets to designated filesystems to prevent mismatches and timeouts.
Exporting SLURM_CONF before running squeue, sinfo, or sbatch is required on SONIC to ensure the SLURM commands correctly target the cluster configuration, preventing command failures and job submission errors.
Managing CUDA versions for gorgonas and medusas requires selecting compatible environments matching cu124, cu128, or cu130, while properly configuring venv paths to avoid runtime mismatches during GPU-accelerated training runs.
To minimize I/O bottlenecks on SONIC, stage datasets to fast filesystems such as /tmp, /snfs2, or /sonic_home, utilizing dataset caching strategies to optimize data loading during model evaluation and training.
Yes, this workflow supports launching training runs, data processing, and model evaluation across different GPU architectures like gorgonas and medusas by guiding partition selection and environment variable management.
Limitations when deploying on SONIC include potential dataset staging mismatches if not placed on /tmp, /snfs2, or /sonic_home, and job timeouts if idle partitions are not strictly enforced before submission.
sayalinvidiaConfigure NeMo AutoModel launches across interactive, Slurm, and SkyPilot environments.
jakechen1Submit, monitor, and manage Slurm jobs on the Cheaha HPC cluster.
randoneeringProvision pinned NVIDIA CUDA toolkits and libraries in Linux FLOX environments.
umich-foreseerGenerate Slurm sbatch scripts with partition, account, GPU, and validation directives.
