Dispatch SLURM workloads by wrapping srun inside the current allocation.
npx skills add https://github.com/mlc-ai/pith-train --skill launch-with-slurm
Please help me install this Agent Skill. Skill: launch-with-slurm Source: https://github.com/mlc-ai/pith-train/tree/main/.agents/skills/launch-with-slurm Command: npx skills add https://github.com/mlc-ai/pith-train --skill launch-with-slurm
Dispatching work inside a SLURM allocation using srun to efficiently utilize reserved compute on single-node or multi-node clusters.
Invoke srun with the appropriate flags inside the current SLURM allocation to start your job.
High-intent search queries and answers about installing and using this skill.
To launch a training job inside a SLURM allocation, wrap srun to dispatch the workload while preserving environment and IO behavior across ranks. The skill reads allocation context with scontrol and builds an srun command using -N, --nodelist, and -o to execute properly.
Use srun instead of raw torchrun when $SLURM_JOB_ID is set to preserve environment and IO behavior across ranks. Using srun ensures context-aware dispatch within the active allocation, properly utilizing reserved compute on single-node or multi-node clusters.
Scontrol reads AllocTRES and NodeList to tailor srun commands to the specific allocation. This context-aware dispatch ensures the built srun command uses the correct -N flag and --nodelist parameter to target reserved compute resources across multi-node clusters.
Yes, you can launch multi-node workloads across existing SLURM allocations. The method applies to both single-node and multi-node allocations, reading the allocation context via scontrol to build appropriate srun commands that target the correct nodes.
To capture distributed logs and support resumed runs, use the -o flag combined with --open-mode in the srun command. This enforces proper logging requirements and ensures output is captured correctly for distributed training jobs inside the allocation.
Using srun inside a SLURM allocation requires an active allocation context with $SLURM_JOB_ID set, and it enforces strict executable requirements. It cannot dispatch work outside of an existing allocation and relies on scontrol to read AllocTRES and NodeList correctly.
yo-stevenConvert single-node PyTorch distributed launches into multi-node Slurm sbatch jobs.
crim-caExtracts job execution logs with I/O streams and timestamps from Weaver API-driven systems.
TianyuDuDiscover Slurm cluster configurations and validate job submissions via sbatch, salloc, and srun.
ovachieverProcess distributed ML data pipelines with Ray Data across CPU/GPU clusters.
