run-experiment

To run ML experiments on remote GPUs, this Skill detects your environment via CLAUDE.md, synchronizes code to SSH servers, Vast.ai, or Modal, and launches jobs automatically. It performs pre-flight GPU checks and uses screen sessions for execution.

Yes, you can deploy to Vast.ai and control costs by enabling auto-destroy. The Skill provisions instances, launches training, collects results and logs, and tears down the instance to stop billing automatically upon completion.

Modal serverless is fully supported for launching machine learning training scripts. When Modal is detected in your configuration, the Skill delegates to a serverless-modal launcher to orchestrate the deployment and execution of your jobs.

The best way to sync code and log metrics for remote GPU training is using built-in rsync or git/pull workflows. The Skill synchronizes necessary files and optionally auto-instruments your training scripts for Weights & Biases (W&B) logging.

To verify a remote ML training job started successfully, the Skill provides specific verification commands after launch. It initiates runs using screen sessions or background processes and supplies the exact commands needed to check job status.

No, you do not need to manually check VRAM availability before deploying to local GPUs. The Skill performs pre-flight GPU validation automatically, querying free VRAM thresholds to ensure the local GPU meets requirements before starting the run.