SkyPilot — мультиоблачный запуск ML-задач

★ 7.7 · orchestration

skypilot is a Claude Code skill that enables multi-cloud orchestration of ML workloads with automatic cost optimization using SkyPilot. It covers launching training and batch jobs across 20+ providers — AWS, GCP, Azure, Kubernetes, RunPod, and more — with automatic selection of the cheapest available cloud and region. Spot instance support delivers 3–6× cost savings with built-in auto-recovery on preemption, while distributed multi-node training relies on gang scheduling and environment variables such as SKYPILOT_NODE_RANK, SKYPILOT_NUM_NODES, and SKYPILOT_NUM_GPUS_PER_NODE. Core commands include sky launch, sky exec, sky jobs launch, and sky serve up; managed jobs support checkpointing and fault tolerance for long-running workloads. The skill is aimed at ML engineers who need to cut GPU costs and avoid vendor lock-in without being tied to a single cloud provider.