gpu-cli_skill

Overview

This skill provides a CLI that runs local commands on remote NVIDIA GPUs (A100, H100, RTX 4090) by prefixing them with 'gpu'. It provisions pods, syncs project files, streams logs, and syncs outputs back so you can run training, LLM inference, ComfyUI workflows, and notebooks transparently. The tool supports interactive and detached jobs, volume management, encrypted outputs, and JSON output for automation.

How this skill works

Prefix any shell command with 'gpu' and the CLI provisions a remote pod with the requested GPU, syncs your code and dependencies, runs the command remotely, streams logs, and pulls outputs back to your workspace. It reuses pods to avoid repeated provisioning, supports detached jobs, and exposes machine-readable --json output for status, logs, inventory and other commands. Configuration lives in project files or global config and secrets use the OS keychain.

When to use it

• Run ML training on remote NVIDIA GPUs without changing local workflows (e.g., gpu run python train.py).
• Launch LLM inference or model serving with vLLM or Ollama via gpu llm run.
• Execute ComfyUI workflows or text-to-image generation on powerful GPUs.
• Run notebooks or automated CI jobs that need GPU acceleration with environment sync.
• Create persistent volumes, migrate storage, or sync outputs across datacenters.
• Automate monitoring and billing with --json output for scripts and dashboards.

Best practices

• Use gpu login and gpu init before running jobs to ensure authentication and project config.
• Control what is uploaded with .gitignore and outputs patterns to avoid transferring secrets or large irrelevant files.
• Prefer pod reuse for short iterative runs to reduce provisioning overhead; gpu stop to release resources when finished.
• Use volumes for large datasets or model caches to avoid repeated uploads.
• Rely on --json on CI or tooling to parse status, logs, inventory and reduce flaky parsing of human output.

Example use cases

• Train a PyTorch model remotely: gpu run python train.py --epochs 10 --gpu-type "A100"
• Run detached long training and reattach later: gpu run -d python long_run.py then gpu attach <job_id>.
• Launch an LLM endpoint for inference tests: gpu llm run --vllm --model mymodel -y and monitor with gpu llm info.
• Execute a ComfyUI workflow for image generation: gpu comfyui run my_workflow and download outputs via gpu vault export.
• Run CI GPU tests using env vars (GPU_RUNPOD_API_KEY) and parse gpu status --json in the pipeline.