dstack

dstack is the open-source control plane for GPU orchestration: it unifies provisioning and management for AI development, training, and inference.

This platform delivers a unified control plane for ML teams, orchestrating GPU resources across multiple environments: cloud, Kubernetes, or on-prem. It acts as a streamlined alternative to complex systems like Kubernetes or Slurm, specifically engineered for AI workloads. Users define tasks (dev environments, training, services) via simple configuration, and dstack handles provisioning, scheduling, and auto-scaling. The result: teams report reducing infrastructure costs by 3-7x while maintaining compatibility with diverse hardware (NVIDIA, AMD, TPUs) and open-source tools.

https://dstack.ai