Dstack

dstack is an open source control plane for GPU provisioning and orchestration of AI workloads. It gives one interface for development, training and inference across GPU clouds, Kubernetes, VMs and bare metal servers, and handles infrastructure provisioning, job scheduling, auto scaling, port forwarding and ingress so teams do not need to build their own AI compute stack. It is aimed at ML engineers, AI labs, cloud tenants and data center operators. Work is described in YAML configurations kept in the repository: fleets provision and manage clusters, dev environments open machines for an IDE or for agents, tasks run training or batch jobs on one node or many, services deploy model inference as secure scalable endpoints, and volumes persist data. An experimental Presets feature applies agent driven inference optimization. Configurations are applied through the dstack apply CLI command or a programmatic HTTP API. It supports NVIDIA, AMD, TPU and Tenstorrent accelerators, integrates natively with major GPU clouds using the customer's own credentials, connects to existing hosts over SSH, and can run on top of Kubernetes or, through an experimental backend, Slurm clusters. The core dstack server is free to self host and published on GitHub under the MPL 2.0 license. dstack Sky, the hosted option, offers unified access to GPU clouds with one bill, and dstack Factory is a multi tenant stack for labs and data centers; neither has published plan prices. dstack positions itself as a lighter alternative to Kubernetes and Slurm for AI work, while still being able to run on top of either.