What problem does it solve? Launching ML experiments on GPU machines involves repetitive manual steps: checking GPU availability, syncing code, setting up screen sessions, and verifying processes started correctly. This Skill automates that deployment workflow so experiments launch consistently on the right GPU with logs captured. ## Core Features & Use Cases - Environment Detection: Reads the project's CLAUDE.md to determine whether to run locally (CUDA/MPS) or on a remote server via SSH, including conda environment and code directory settings. - GPU Pre-flight Checks: Queries nvidia-smi (or PyTorch MPS on Mac) to find free GPUs before assigning any job, preventing memory conflicts. - Code Sync & Deployment: Syncs code via rsync or git push/pull, then launches each experiment in a dedicated screen session (remote) or background process (local) with CUDA_VISIBLE_DEVICES binding and tee-based logging. - Use Case: You have three training runs to launch on a remote 4x A100 server. The Skill checks which GPUs are free, rsyncs your Python files, starts three screen sessions each pinned to a different GPU, and reports back the GPU assignments and log file locations. ## Quick Start Ask the AI to run your training experiment on the GPU server, for example by saying run experiment: train the NVRC codec model on the remote server.
