sue-interface-check

Validating ML experiment code for scale-up readiness requires auditing the codebase against 22 prioritized rules covering dry-run profiling, inode-safe output controls, and preflight smoke tests. This ensures code is portable and safe for HPC and cloud backends like LUMI or RunPod before allocating compute.

A dry-run profiling check is a mandatory hard gate that verifies training pipelines execute correctly on minimal resources before full scale-up. It ensures PyTorch DDP support and Hydra configuration function properly, preventing wasted GPU compute on broken or non-portable code.

Enforcing W&B and CSV logging contracts involves auditing the experiment codebase to verify it exposes minimum portable interfaces for queryable progress monitoring. This checks that result ledgers and logging pipelines meet the required standards for safe, cost-effective scale-up.

Yes, the audit evaluates codebases targeting HPC and cloud sandbox backends including LUMI, Snellius, NM5, Brev, RunPod, and AutoDL. It verifies that training, generation, and evaluation pipelines meet the portable interface requirements specific to these platforms.

The four mandatory hard gates for pre-submission validation are real dry-run profiling, inode-saving output mode, first-run smoke test, and a committed running README under scale_up_scripts/. These gates ensure student code is portable and will not waste expensive compute resources.

An inode-safe output mode is required to prevent file system exhaustion during large-scale ML training runs on HPC systems. It acts as a mandatory hard gate during code audits to ensure output controls are properly configured before allocating expensive GPU resources.