single-gpu-dryrun

What problem does it solve?

Running full ML production experiments without prior validation often leads to wasted GPU hours, failed runs due to missing dependencies or configuration errors, and unexpected inode consumption on shared HPC clusters. This Skill eliminates that risk by enabling small, real single-GPU probe runs that catch issues early.

Core Features & Use Cases

• LUMI dev-g Pre-Production Probe: Runs tiny, real experiments on LUMI's dev-g partition to validate environment setup, model construction, and basic training steps before launching expensive full runs.
• Accurate Parameter Counting: Tallies total, trainable, and non-trainable model parameters using real tensor operations, avoiding unreliable estimates from config files or research papers.
• Production Gating Validation: Includes mandatory inode consumption risk checks and minimal evaluation path testing (e.g., FID for generative models) to ensure probes are valid gates for full production launches.
• Use Case: If you are preparing a large-scale image generation experiment on LUMI, use this Skill to run a 1-step dry run first to catch missing dependencies, config errors, or inode shortages before wasting hundreds of GPU hours on a failed full run.

Quick Start

Use the single-gpu-dryrun skill to run a validation probe for your LUMI ML experiment before launching the full production run.