GPU CLI

Run any code on cloud GPUs with a single command. Just prefix your normal commands with gpu run.

python train.py           # local
gpu run python train.py   # remote GPU

• Simple - Prefix commands with gpu run, that's it
• Fast - Connection pooling, delta sync, real-time output streaming
• Cost-efficient - Auto-stops pods when idle (save 60-98% on GPU costs)
• Multi-cloud - RunPod, Vast.ai, local Docker
• Secure - Zero-trust encryption on supported providers

# 1. Install GPU CLI
curl -fsSL https://gpu-cli.sh/install.sh | sh

# 2. Run your code on a remote GPU
gpu run python train.py

This repo includes a Claude Code plugin that supercharges GPU CLI with AI assistance. Describe what you want in plain English, and Claude generates complete, runnable GPU workflows.

Create a LoRA training project:

You: I want to train a LoRA on photos of my dog so I can generate images of it

Claude: [Generates complete project with gpu.jsonc, train.py, requirements.txt, README.md]

Set up a private LLM API:

You: Set up Llama 3.1 70B as a private ChatGPT-like API

Claude: [Generates vLLM server config with OpenAI-compatible endpoints]

Debug an error:

You: /gpu-cli:gpu-debug CUDA out of memory when running FLUX

Claude: [Analyzes error, suggests reducing batch size or upgrading to A100]

Optimize costs:

You: /gpu-cli:gpu-optimize

Claude: [Reviews gpu.jsonc, suggests RTX 4090 instead of A100 for your workload, saving 75%]

Ready-to-use templates for common AI/ML workflows:

# Run a command on remote GPU
gpu run python script.py

# Run a server with port forwarding
gpu run -p 8188:8188 python server.py --listen 0.0.0.0

# Open a shell on the remote pod
gpu shell

# View running pods
gpu pods list

# Stop a pod
gpu stop

# Interactive dashboard
gpu dashboard

Create a gpu.jsonc in your project:

{
  "$schema": "https://gpu-cli.sh/schema/v1/gpu.json",
  "project_id": "my-project",
  "provider": "runpod",

  // Sync outputs back to local machine
  "outputs": ["output/", "models/"],

  // GPU selection
  "gpu_type": "RTX 4090",
  "min_vram": 24,

  // Optional: Pre-download models
  "download": [
    { "strategy": "hf", "source": "black-forest-labs/FLUX.1-dev", "allow": "*.safetensors" }
  ],

  "environment": {
    "base_image": "runpod/pytorch:2.4.0-py3.11-cuda12.4.1-devel-ubuntu22.04"
  }
}

For faster startup and persistent model storage, use RunPod Network Volumes. See the Network Volumes Guide for setup instructions.

• Network Volumes Guide - Persistent storage for models
• GPU CLI Docs - Full documentation

MIT