Dstack

Dstack is an open-source control plane designed to run development, training, and inference workloads on GPUs across heterogeneous infrastructure. It abstracts away the complexity of managing compute on hyperscalers, emerging “neocloud” providers, and on-premise clusters, allowing teams to define and execute jobs through a unified interface. Its primary purpose is to give ML and engineering teams consistent, reproducible environments and efficient GPU utilization regardless of where the hardware resides.

Dstack provides declarative configuration of jobs and environments, so users can define compute requirements, dependencies, and artifacts as code. It integrates with popular cloud providers and Kubernetes-based setups, enabling automatic provisioning, scaling, and scheduling of GPU instances. The platform supports interactive development (e.g., notebooks, IDE servers), batch training jobs, and high-throughput inference services, all managed through a single control plane. Built-in features such as job queuing, resource pooling, and cost-aware scheduling help optimize GPU usage and reduce idle time.