gpu-cli

What This Skill Does

The gpu-cli skill provides a powerful interface for offloading intensive computational workloads from your local environment to high-performance cloud-based NVIDIA GPUs. By acting as a transparent bridge, it allows you to prefix standard local CLI commands with gpu, orchestrating the provisioning of compute pods, intelligent file synchronization of your codebase, and the streaming of remote output back to your terminal. It supports enterprise-grade hardware including A100, H100, and RTX 4090 configurations, making it an essential tool for developers working on ML model training, large language model (LLM) fine-tuning, and heavy-duty rendering workflows like ComfyUI.

Installation

To integrate this capability into your OpenClaw agent, use the following installation command: clawhub install openclaw/skills/skills/angusbezzina/gpu-cli

Use Cases

This skill is designed for scenarios where your local hardware is insufficient for the task at hand. Common use cases include:

• Scaling ML training jobs by leveraging multi-GPU clusters (up to 8 units).
• Running distributed inference tasks that require low-latency access to high-VRAM NVIDIA GPUs.
• Managing complex media generation pipelines that would otherwise stall a local development machine.
• Persistent background compute: running long-duration training sessions via detached modes while maintaining the ability to reattach and inspect logs at any time.

Example Prompts

1. "gpu run python train_model.py --epochs 50 --batch-size 128 using an H100."
2. "Check our current active GPU pods and their running costs to ensure we are within budget."
3. "List all available GPU inventory with at least 24GB of VRAM and filter by the lowest hourly price."

Tips & Limitations

Always initialize your project with gpu init before running your first command to ensure proper configuration. Utilize gpu doctor frequently to verify that your authentication tokens and provider keys are correctly configured. Remember that gpu run automatically synchronizes your local directory; ensure you have a .gitignore set up to prevent unnecessary large files from being synced. The skill supports interactive mode (-i), which is excellent for debugging remote environments in real-time, but be mindful of usage costs, as persistent interactive sessions consume compute resources continuously.