Introduction
GPU-Action (SpotWarp) is a lightweight, 100% local Python autopilot daemon designed to protect your AI training and inference workloads running on cheap Spot instances (Vast.ai, RunPod, AWS Spot) from sudden evictions, ensuring zero data loss and continuous service uptime.
By exploiting Spot pricing, you can save up to 70% on cloud compute bills. When an eviction warning fires, the SpotWarp daemon rents a matching alternative host, migrates your workspace files using high-speed incremental rsync, restarts your training scripts, and binds active endpoints — automatically, without manual intervention.
Your master API credentials (Vast/RunPod keys) and workspace training files remain strictly on your local computer. Rented remote containers never see your private access keys, guaranteeing a true Zero-Trust security model.
Installation
If Python 3.8+ is installed on your local host machine, you can install the official package via Pip in seconds:
pip install spotwarpWindows Rsync Auto-Setup
GPU-Action utilizes rsync for high-speed workspace delta backups. If rsync is not natively installed on your Windows machine, the daemon automatically downloads a portable, sandboxed cygwin-rsync binary to ~/.gpu_action/bin on its first startup.
Environment Configuration
Before starting the guard, export your cloud provider API credentials locally on your client machine:
macOS / Linux
export VAST_API_KEY="your_vast_api_key_here"
export RUNPOD_API_KEY="your_runpod_api_key_here"Windows PowerShell
$env:VAST_API_KEY="your_vast_api_key_here"
$env:RUNPOD_API_KEY="your_runpod_api_key_here"Start Guard Daemon
Start the failover guard by passing your subscription license key:
spotwarp start --license-key YOUR_LICENSE_KEY
For local sandbox evaluation, you can bypass central licensing by passing the debug key DEVELOPMENT_TEST_KEY as the license argument.
Stateful Auto-Resume
You can configure the daemon to automatically resume PyTorch/TensorFlow training scripts on the replacement node pointing to your synced checkpoints:
spotwarp start --license-key YOUR_LICENSE_KEY --resume-cmd "python /workspace/train.py --resume"
When eviction occurs, GPU-Action syncs the final state of your workspace directory, boots the cheapest matching host, restores all cached weights, and executes the --resume-cmd in the background via remote SSH/nohup.
Console Dashboard Sync
Once the daemon starts, it registers active telemetry with the central server:
- Log in to the Live Console Dashboard using your license key.
- Monitor active instances (specs, uptime, SSH ports), cumulative cost savings counter ($ / %), and live daemon status logs in real time.
Security & Sandbox
All API calls and status heartbeat transmissions are encrypted in transit via HTTPS using AES-256 tokens.
The CLI daemon runs strictly in user space on your local computer. It does not require root/sudo access, and can run inside isolated docker containers, making it completely safe for enterprise production environments.