개요
dstack provides a unified interface for managing compute across cloud backends, Kubernetes clusters, and bare-metal or on-premises servers. Repository-based YAML configurations describe fleets, development environments, tasks, services, and persistent volumes. Configurations can be applied through the CLI, a programmatic API, or supported AI agent skills. The repository metadata identifies Python as the primary language and shows that the project is active rather than archived.
주요 기능
- Vendor-agnostic orchestration across GPU clouds, Kubernetes, and on-premises clusters
- Support for NVIDIA, AMD, Google TPU, and Tenstorrent accelerators
- YAML configurations for fleets, development environments, tasks, services, and volumes
- CLI, programmatic API, and AI agent skill workflows
- Distributed jobs, model deployment, web applications, autoscaling, and authorization
- Automated provisioning, job queuing, networking, volume handling, port forwarding, and failure handling
- SSH fleets for on-premises servers without backend configuration
요구 사항, 설치 및 빠른 시작
사용 정보
모델 호환성 및 사용 사례
The repository describes compatibility with open-source tools and frameworks but does not list specific model families. It supports inference and model deployment on NVIDIA, AMD, Google TPU, and Tenstorrent hardware. Claude, Codex, and Cursor are mentioned as agents that can use dstack skills; they are not identified as hosted inference models. MCP and RAG support are not stated in the repository metadata.
라이선스 및 위험 참고 사항
The repository metadata specifies the Mozilla Public License 2.0 (MPL-2.0). Review the repository license before redistribution or modification: https://api.github.com/licenses/mpl-2.0
릴리스 및 유지 관리
The README lists releases including 0.20.17 for PD disaggregation and Kubernetes volumes, 0.20.16 for performance and an SSH proxy, 0.20.13 for exports and templates, 0.20.12 for Crusoe, 0.20.8 for skills, and 0.20.0 for a fleet-first UX and events. Full details are available from the repository releases linked in the README.