dstack
Open-source control plane for GPU provisioning and orchestration across clouds, Kubernetes, and on-prem clusters. Simplifies AI workload management with a container-first approach.
dstack is an open-source alternative to:
dstack is an open-source orchestration layer designed for heterogeneous AI compute. It standardizes workload execution across GPU clouds, Kubernetes, and on-prem clusters, supporting NVIDIA, AMD, Tenstorrent, and TPU accelerators.
Key Features
- Vendor-agnostic orchestration: Unify fleets, dev environments, tasks, services, volumes, and gateways in one control plane for AI workloads
- Cloud & on-prem support: Deploy across AWS, GCP, Azure, Kubernetes, bare-metal servers, or VMs with SSH access
- AI-native primitives: First-class support for accelerator provisioning, workload scheduling, and observability tailored for ML/AI workloads
- Marketplace access: Leverage dstack Sky for hosted GPU provisioning without managing cloud accounts
- Cost optimization: Smart orchestration with fine-grained fleet controls, including spot, on-demand, and reserved capacity
- Simple interfaces: Lightweight CLI, UI, and API that eliminate Kubernetes/Slurm complexity
- Framework agnostic: Works with any open-source AI frameworks and tools, Docker images, and custom code
Tech / Self-Hosting
- Open-source core: Self-hostable server with YAML-based configuration for fleets, backends, and projects
- Backend integrations: Native support for major GPU clouds (AWS, GCP, Lambda, Nebius, Crusoe, RunPod, Azure, Verda, Digital Ocean, JarvisLabs, CloudRift) and Kubernetes clusters
- Deployment options: Run as a standalone server or integrate with existing infrastructure
- Configuration examples:
```yaml projects:
- name: main
backends:
- type: aws
creds: type: accesskey accesskey: KKAAUKLIZ5EHKICAOASV secret_key: pn158lMqSBJiySwpQ9ubwmI6VUU3/W2fdJdFwfgO ```
- Interoperability: 100% compatible with existing stacks and tooling