dstack

What it is and what it does

dstack is an orchestration engine that acts as a central control plane for GPU workloads, abstracting away the complexity of provisioning and managing compute across multiple cloud providers, Kubernetes clusters, and on-premises infrastructure. It handles provisioning, job queuing, auto-scaling, networking, volumes, and failure recovery automatically. The package includes a server component that coordinates resources and a CLI tool for submitting workloads. Users define configurations in YAML for fleets (clusters), dev environments, tasks (training or batch jobs), services (inference endpoints), presets (optimization), and volumes, then apply them via the CLI or programmatic API.

The system supports multiple accelerator types out of the box and integrates with popular ML frameworks. It requires a running dstack server (which can be deployed on Linux, macOS via WSL 2, or Windows) and Git and OpenSSH as system dependencies. The package has 25 runtime dependencies including pydantic, requests, paramiko, and rich, providing configuration validation, HTTP communication, SSH connectivity, and CLI formatting.

Use it for:

• Provision and manage GPU clusters across AWS, GCP, Azure, and on-premises without rewriting infrastructure code for each provider
• Launch development environments on remote GPU hardware for interactive work or IDE integration
• Submit distributed training jobs that auto-scale across multiple nodes and handle out-of-capacity errors automatically
• Deploy model inference services as secure, scalable endpoints with replica management and load balancing
• Manage persistent storage volumes across cloud and on-premises infrastructure for data sharing between workloads

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

dstack is a unified control plane for provisioning and orchestrating GPU workloads across cloud providers, Kubernetes, and on-premises clusters, supporting NVIDIA, AMD, Google TPU, and Tenstorrent accelerators.

Yes, if you need to orchestrate GPU workloads across multiple cloud providers or Kubernetes clusters. The active maintenance, low install friction, recent updates (including Pydantic v2 support), and zero known vulnerabilities make it a solid choice. The copyleft license requires compliance with MPL 2.0 terms if you distribute or modify the code. Requires Python 3.10+ and system-level Git and OpenSSH setup.

Install

dstack on PyPI

pip install dstack

uv add dstack

poetry add dstack

Installing dstack

Before you install

Low install friction with a pure-Python wheel distribution. The package is actively maintained with a recent release (0.21.1) and 2213 GitHub stars. Requires Python 3.10 or later and Git and OpenSSH as system dependencies.

License in practice

Licensed under Mozilla Public License 2.0 (copyleft). Users must comply with MPL 2.0 terms when distributing or modifying dstack; derivative works must be made available under the same license.

Quickstart

# Install dstack
$ uv tool install "dstack[all]" -U

# Start the server
$ dstack server

# In another terminal, configure the CLI
$ dstack project add --name main --url http://127.0.0.1:3000 --token <admin-token>

# Use the CLI to apply configurations
$ dstack apply -f fleet.yaml

Requires Git and OpenSSH to be installed on the system. Server requires Python 3.10+. Backend configuration is needed for cloud or Kubernetes orchestration; on-premises use requires SSH fleet setup.

Verify before relying

• Whether the package works with Python 3.11+ or only 3.10 specifically
• Performance characteristics and scalability limits for large clusters
• Detailed cost estimation or billing integration capabilities
• Support status for each accelerator type (NVIDIA, AMD, TPU, Tenstorrent)

Package facts

Evidence: dstack-0.21.1-py3-none-any.whl

Tags

More Artificial Intelligence packages

Further reading