sue-parallelism-discussion

What problem does it solve?

This Skill eliminates wasted GPU resources, failed job submissions, and inconsistent experiment results caused by unplanned parallelism, poor node packing, and unvalidated concurrency strategies for ML/HPC scale-up experiments run on remote sandbox backends like LUMI, Snellius, and RunPod.

Core Features & Use Cases

• Pre-submission parallelism gate: Acts as a mandatory deliberation step between launcher script writing and dryrun submission, ensuring all parallelism plans are explicit and auditable before any compute resources are consumed.
• GPU-occupancy-first planning: Enforces rules that pack full nodes with exactly one worker per GPU, only allowing partial final nodes, to maximize hardware utilization and avoid idle GPU waste.
• Contract validation and visualization: Integrates with SUE family audit tools and Gantt chart generation to validate plans against family-wide contracts and visualize wave scheduling, node allocation, and potential idle GPU periods.
• Use Case: A researcher preparing to run 20 generation variants on a LUMI sandbox can use this Skill to calculate the optimal max_parallel value, node packing shape, and wave schedule that avoids queue limit failures and ensures all allocated GPUs are used efficiently.

Quick Start

Use the sue-parallelism-discussion skill to create a validated, auditable parallelism plan for your DeepResearch scale-up experiment after launcher scripts are written and before submitting any dryrun or fullrun jobs to your selected sandbox backend.