sue-dryrun

What problem does it solve?

This Skill eliminates the risk of wasting expensive GPU compute on full ML/HPC scale-up experiments that fail due to uncaught script errors, environment mismatches, or path non-compliance, by running a minimal representative dry run to validate all critical paths before full launch.

Core Features & Use Cases

• Plan-only matrix validation: Enumerates the full experiment matrix, counts, uniqueness, and launch commands without executing every variant, catching configuration errors early.
• Minimal runtime coverage: Runs a small set of representative real cases covering all distinct evaluator, data-loader, asset, scheduler, tracking, and output paths that could fail independently in the full run.
• Runtime and makespan estimation: Measures dryrun throughput to estimate total full-run runtime and sandbox completion time under maximum GPU quota and job concurrency.
• Use Case: For a 500+ variant image classification scale-up experiment, use this skill to run 5-10 representative dry runs to confirm all code and data paths work, estimate total runtime, and catch LUMI container environment mismatches before launching the full expensive grid.

Quick Start

Use the sue-dryrun skill to run a minimal readiness dry run for your DeepResearch scale-up experiment after your launch scripts are written and audited.