dstack

dstack is an open-source orchestration platform built for AI teams. It manages the complex task of provisioning GPU resources and running containerized AI workloads. You can use it to get GPUs from various cloud providers, your own Kubernetes clusters, or even bare-metal servers, all through a single interface. Its main job is to let developers and researchers run their code—whether it's for development, training, or serving models—without getting bogged down by infrastructure details.

The platform works by providing a unified control plane. You define what you need—like a development environment, a training task, or an inference service—using simple configuration files. dstack then handles the rest: it finds available GPUs, spins up the necessary instances, manages the container lifecycle, and can even orchestrate multi-node clusters for distributed workloads. It integrates natively with cloud APIs for speed and supports open-source frameworks, keeping your workflows portable and reproducible.

This tool is built for AI teams who need efficient, flexible access to computing power. Machine learning engineers can use it to quickly get a GPU-powered coding environment connected to their IDE. Research teams can use it to launch large-scale training jobs across multiple nodes. Companies deploying models can use it to host scalable, OpenAI-compatible inference endpoints. It helps these groups increase GPU utilization, control costs, and avoid being locked into a single cloud vendor.