What problem does it solve? Running GPU workloads like model training, fine-tuning, and inference typically requires SSH access, Docker setup, and manual instance management. This Skill generates and runs Modal launcher scripts so you can execute GPU jobs from your laptop with per-second billing and automatic scale-to-zero. ## Core Features & Use Cases - Cost Estimation Before Every Run: Calculates VRAM requirements from model size, selects the right GPU (T4 through B200), and presents a cost estimate for confirmation before execution. - Six Launcher Patterns: Generates Modal scripts for one-shot training, persistent web API inference, vLLM serving, batch parallel processing, LoRA fine-tuning, and multi-GPU distributed training. - Result Collection & Cleanup: Retrieves outputs from Modal Volumes or stdout and stops deployed apps, with no manual instance teardown since billing stops automatically. - Use Case: You need to benchmark a 7B model but have no local GPU. The Skill estimates that an H100 run costs about $0.98 for 15 minutes, generates the launcher script, runs it with modal run, and returns results locally. ## Quick Start Ask the assistant to run your training script on Modal with an A100 GPU and confirm the cost estimate before it executes.
