What problem does it solve? Renting cloud GPUs for ML experiments requires manually estimating VRAM needs, comparing offers, configuring SSH access, and remembering to tear down instances to stop billing. This Skill automates the full vast.ai instance lifecycle so you describe the training task and get a cost-optimized GPU without picking hardware yourself. ## Core Features & Use Cases - Task-driven GPU selection: Analyzes experiment plans and scripts to estimate VRAM, GPU count, disk, and training hours, then presents three offers ranked by estimated total cost. - Full lifecycle management: Handles rent, environment setup (phased pip installs, code sync via rsync), result download, and instance destruction with cost reporting. - State tracking: Maintains vast-instances.json so other skills like /run-experiment and /monitor-experiment can connect to active instances. - Use Case: You need to fine-tune a 7B model for about 4 hours. The skill estimates ~16 GB VRAM, searches vast.ai offers, shows options like an RTX 4090 at ~$1.12 total, rents it, syncs your code, and destroys the instance when training finishes. ## Quick Start Ask the assistant to rent a GPU on vast.ai for your training task, for example: rent a vast.ai GPU to fine-tune my LLaMA-7B model for about four hours.
