SkyPilot: Run AI on Any Cloud

SkyPilot is an open-source framework for running LLMs, AI, and batch jobs on any cloud. In this talk, we describe how SkyPilot offers AI practitioners high GPU availability, maximum cost savings, and managed execution — all with a simple cloud-agnostic interface. SkyPilot is open source (Apache 2.0) and under active development at UC Berkeley.

SkyPilot's new approach is to view all clouds and regions (that a user has access to) as a coherent collection called the Sky. SkyPilot transparently sends a job to the cheapest and most available location in the Sky and manages its execution. It cuts cloud costs by supporting spot instances with automatic recovery (potentially across zones, regions, clouds) and automatically cleaning up idle cloud instances.

We'll also share our experiences deploying SkyPilot to users in the past year. SkyPilot has powered the development of popular LLM projects including the Vicuna chatbots and the vLLM inference system. It is being used by 10s of organizations for diverse use cases, including AI training on GPUs/TPUs, AI serving, CPU batch jobs, and interactive development.