SkyPilot
Open-source AI compute management software that runs jobs across your Kubernetes and Slurm clusters or cloud accounts, using GPUs, TPUs and CPUs.
SkyPilot is an open-source system for AI teams that need to run training, inference and development workloads across their own clusters and cloud accounts. It brings Kubernetes, Slurm and cloud compute under one interface, so teams can move jobs between providers without rewriting their workload code.
Compute runs within your accounts, VPCs and clusters. SkyPilot can use existing GPU resources or provision GPUs, TPUs and CPUs through providers including AWS, GCP, Azure, CoreWeave and RunPod. Its portable job definitions describe the environment and workload together, reducing the work involved in moving a job to different infrastructure.
SkyPilot handles job queues, automatic recovery and failover when a provider lacks capacity. For distributed workloads, it supports multi-node jobs and gang scheduling, which schedules a job's required resources together. A shared control plane lets infrastructure teams manage multiple clusters and allocate resources across a team.
It also addresses idle and underused compute. Automatic stopping cleans up idle resources, while workload packing fits jobs onto shared clusters. The scheduler considers availability and cost when choosing where a job should run.
Developers can connect an IDE to Kubernetes pods, sync code and use SSH for interactive work. SkyPilot supports existing GPU, TPU and CPU workloads without code changes and is licensed under Apache 2.0.
