GPU Orchestration Platform
2.1k 2026-04-18
dstackai/dstack
A vendor-agnostic unified control plane for GPU provisioning and orchestration across clouds, Kubernetes, and on-prem for AI/ML workloads.
Core Features
Vendor-agnostic GPU orchestration (NVIDIA, AMD, TPU, Tenstorrent).
Supports multi-environment deployment (clouds, Kubernetes, bare metal).
Streamlines development, training, and inference workflows.
Configuration via YAML for fleets, dev environments, tasks, services, and volumes.
Integration with AI agents for automated management.
Quick Start
uv tool install "dstack[all]" -UDetailed Introduction
dstack is an open-source, unified control plane designed to simplify GPU provisioning and orchestration for AI/ML workloads. It offers vendor-agnostic support for various accelerators (NVIDIA, AMD, Google TPU, Tenstorrent) and deployment environments, including major cloud providers, Kubernetes clusters, and on-premise infrastructure. By providing a consistent interface for managing compute resources, dstack streamlines the entire MLOps lifecycle, from interactive development and model training to inference and service deployment, making it easier for teams to build and scale AI applications efficiently.