Modal — запуск ML на serverless GPU

★ 7.7 · devops

modal is a Claude Code skill that enables running ML workloads on Modal's serverless GPU cloud without any infrastructure management. It covers GPUs ranging from T4 (16 GB) to H200 (141 GB) and B200 (Blackwell), with per-second billing and instant auto-scaling from zero to 100+ GPUs. Infrastructure is defined in pure Python — no YAML — using Modal primitives: App, Function, Cls, Image, Volume, and Secret, with modal run, modal serve, and modal deploy commands covering the full development-to-production cycle. The skill includes deploying models as REST APIs via FastAPI and ASGI decorators, dynamic batching, container image layer caching for fast iteration, and persistent Volume storage for model caching. It's aimed at ML engineers who need on-demand GPU access for inference, batch training, or scheduled jobs without idle costs or reserved instance commitments.