run-experiment

To deploy and run ML experiments on a remote GPU server, the skill automates SSH connections, syncs necessary code files, verifies GPU availability, and launches the training job in an isolated screen session. It reads target environment configurations from CLAUDE.md to ensure accurate deployment.

Yes, you can run PyTorch training jobs on local machines with CUDA or MPS. The skill automatically detects the local hardware environment, performs pre-flight GPU availability checks, and launches the experiment as a background process with appropriate GPU binding.

Automating MLOps workflows for training jobs requires environment detection, code synchronization, and resource allocation. This skill standardizes the process by reading configuration files, verifying GPU memory, binding resources, and managing processes using screen and conda across local or remote setups.

Code synchronization for remote GPU servers works by efficiently transferring only the necessary code files over SSH. The skill identifies required files locally and syncs them to the remote server before deployment, ensuring the target environment has the up-to-date code needed to run the experiment.

You need conda and screen to manage remote training jobs because the skill uses conda for environment isolation and screen to launch experiments in persistent isolated sessions. This ensures processes continue running after SSH disconnection and provides verification of successful job launches.

ML experiment deployment may fail due to insufficient GPU memory if resources are already allocated. The skill performs pre-flight checks to verify GPU availability and memory before deployment, preventing launch failures and ensuring efficient resource utilization on the target machine.