sue-genai-autotune-hyperparam

What problem does it solve?

This Skill eliminates the manual, error-prone work of testing training hyperparameters for ML/HPC scale-up experiments, preventing OOM failures and wasted GPU hours during full production runs.

Core Features & Use Cases

• Sandbox-First Probe Execution: Runs hyperparameter tests directly on the target production sandbox to ensure results match real-world conditions, avoiding local environment discrepancies.
• Evidence-Backed Recommendations: Generates a recommended_hyperparam.json artifact with GPU telemetry-backed settings that include 10-15% memory headroom to avoid crashes during full runs.
• SUE Pipeline Integration: Sits between launcher script writing and the dryrun step in the SUE workflow, ensuring training experiments only proceed after hyperparameters are validated.
• Use Case: For example, if you are scaling a diffusion model training run on LUMI, this skill automatically tests batch sizes from 1 to 64 to find the largest stable size that fits in GPU memory, skipping manual trial-and-error.

Quick Start

Use the sue-genai-autotune-hyperparam skill to automatically validate and recommend optimal training hyperparameters for your current scale-up experiment on the selected sandbox before proceeding to the dryrun check.