dstack

AI teams use dstack to orchestrate training, inference, and agentic workloads without adopting Kubernetes or Slurm, defining dev environments, tasks, services, and fleets in simple YAML that the Python-based orchestrator provisions anywhere: major clouds, existing Kubernetes clusters, or SSH-reachable bare metal. It is deliberately vendor neutral across accelerators, scheduling onto NVIDIA and AMD GPUs, Google TPUs, and Tenstorrent hardware, and it handles spot instance recovery, service auto-scaling, volumes, and multi-node distributed jobs. The open source server self-hosts under the MPL-2.0 license, while dstack Sky, a hosted version with marketplace GPU pricing, forms the paid tier, so the project is effectively open core. Compared with raw Kubernetes it trades generality for a workflow shaped specifically around ML engineers, which is why infrastructure teams at GPU-heavy startups adopt it; the repository holds around 2,200 GitHub stars with commercial backing from the dstack company. Everything is driven from a CLI and declarative configs, keeping infrastructure reproducible.