About this project

SkyPilot is a system for running, managing and scaling AI workloads on any AI infrastructure. It targets two audiences: AI teams that want a simple interface to launch jobs on any infra, and infrastructure teams that need a unified control plane with scheduling, scaling and orchestration. Core capabilities described in the README: - Unified interface: a task is defined once (YAML or Python API) with resource requirements, workdir, setup commands and run commands, then launched on any available infrastructure. This is presented as a way to avoid vendor lock-in and move jobs between providers. - Multi-infrastructure support: Kubernetes, Slurm, and a long list of clouds and providers including AWS, GCP, Azure, OCI, CoreWeave, Nebius, Lambda Cloud, RunPod, Fluidstack, Cudo, Digital Ocean, Paperspace, Cloudflare, IBM, Vast.ai, VMware vSphere, Seeweb, Prime Intellect, Shadeform, Verda Cloud and Crusoe. - Provisioning and failover: SkyPilot finds available infrastructure, provisions GPUs/TPUs/CPUs, and can automatically fail over when capacity errors occur. - Job management: queueing, running and auto-recovery of many jobs, plus autostop for idle resource cleanup and binpacking on shared clusters. - Kubernetes experience: SSH into pods, sync code, connect an IDE, gang scheduling, multi-node jobs, multi-cluster and multi-cloud control plane. - Team features: deployment via an API server and resource sharing. - Agent integration: a SkyPilot Skill can be installed so coding agents such as Claude Code or Codex can drive SkyPilot. Installation is via uv or pip with extras for selected clouds. The README also points to a quickstart, a YAML spec reference, runnable examples for training (Verl, Llama 4 finetuning, TorchTitan, PyTorch, DeepSpeed, NeMo, Ray, Unsloth, Jax/TPU, OpenRLHF), serving (vLLM, SGLang, Ollama), models (DeepSeek-R1, Llama 4, Llama 3, CodeLlama, Qwen, Kimi-K2, Mixtral), AI applications (RAG, vector databases) and frameworks (Airflow, Jupyter, marimo). The project states it is BYOC: resources are launched within the user's own cloud accounts, VPCs and clusters. Documentation, a blog, Slack, GitHub Discussions and issue tracker are linked for support and contributions.