sue-fullrun-diagnose

To diagnose GPU waste from completed HPC experiments, reduce existing per-shard timing and log artifacts to attribute GPU hours to discrete waste buckets without launching new compute. This fills missing GPU resource utilization metrics in post-run summaries.

GPU hour waste attribution is the forensic analysis of completed ML scale-up runs to assign GPU hours to discrete waste buckets like cancelled retries and setup overhead. It replaces missing GPU-R cells with real values to quantify efficiency gaps.

Yes, you can analyze GPU efficiency breakdowns without launching new compute by applying post-summary forensic analysis to existing full run bundles. This reduces log artifacts to generate per-variant stage timing diagnostics and enriched outputs.

GPU diagnostics for HPC sandboxes requires no specific dependencies or components, operating directly on existing full run bundles. It applies to supported HPC sandboxes where initial run summaries contain missing GPU-R cells.

Use post-run forensic analysis for GPU resource utilization when initial run summaries have missing GPU-R cells, users suspect low GPU efficiency, or per-variant stage timing diagnostics are needed for existing runs on supported HPC sandboxes.