Gpu Cloud Provisioner

The problem

Provisioning GPU instances manually is slow and error-prone. You waste money when you forget to terminate a box after training completes or overpay by not comparing real-time spot prices across vendors.

What it does

• Queries Lambda Labs, RunPod, and Vast.ai APIs to compare real-time price and availability.
• Launches instances with specific images and configures SSH access automatically.
• Synchronizes local project code and datasets to the remote instance via rsync.
• Orchestrates headless execution using tmux to ensure training survives connection drops.
• Monitors GPU health and terminates instances immediately upon job completion to stop billing.

Frameworks & tools

Lambda Labs API, RunPod API, Vast.ai API, SSH, rsync, tmux, and nvidia-smi.

Why this beats prompting it yourself

Generic LLMs don't handle the stateful lifecycle of a cloud instance. This skill manages the specific API sequences, polling logic, and cleanup triggers required to prevent runaway cloud bills.

Use cases

• Fine-tuning large language models on A100 or H100 clusters.
• Running intensive kernel benchmarks on consumer GPUs like the RTX 4090.
• Automating long-running training jobs that require persistent remote sessions.
• Optimizing compute costs by programmatically selecting the cheapest available spot instance.

Known limitations

Requires pre-configured API keys for vendors and pre-uploaded SSH keys in the respective cloud consoles. Not intended for permanent infrastructure or reserved instances.