dstack

Tracked

Vendor-agnostic orchestration for training, inference, and agentic workloads across NVIDIA, AMD, TPU, and Tenstorrent on clouds, Kubernetes, and bare metal.

Author dstackai Open Sourced 2022-01-04 Last Commit Unknown

Introduction

dstack is an open-source orchestration platform for AI development that lets teams run training, fine-tuning, inference, and agentic workloads on any cloud, Kubernetes, or bare-metal infrastructure. It abstracts away compute provisioning so GPU resources from multiple providers and accelerators (NVIDIA, AMD, TPU, Tenstorrent) can be used through a single declarative interface.

Key Features

  • Vendor-agnostic support for clouds, Kubernetes, and on-prem bare metal
  • Unified provisioning across NVIDIA, AMD, TPU, and Tenstorrent accelerators
  • Declarative run and task configuration for training, fine-tuning, and serving
  • Built-in fleet and policy management for cost control and capacity sharing
  • Drop-in acceleration for popular frameworks via containerized environments

Use Cases

  • Provisioning elastic GPU capacity across multiple cloud providers
  • Running distributed training and fine-tuning jobs without infra glue code
  • Deploying and scaling inference services on Kubernetes or bare metal
  • Centralizing compute policy and cost governance for ML teams

Technical Highlights

  • Declarative YAML-based run definitions replace manual provisioning scripts
  • Designed as an open alternative to vendor-specific GPU platforms
  • Integrates with Kubernetes and Slurm-style scheduling workflows
  • Python-native CLI and API for programmatic workload orchestration