DEV.co
AI Frameworks · gpustack

gpustack

GPUStack is an open-source GPU cluster manager that orchestrates inference engines (vLLM, SGLang, TensorRT-LLM) across multi-cloud and on-premises GPU clusters. It provides automated model deployment, load balancing, monitoring, and on-demand SSH-accessible GPU instances for AI workloads.

Source: GitHub — github.com/gpustack/gpustack
5.3k
GitHub stars
565
Forks
Python
Primary language
Apache-2.0
License (OSI-approved)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Repositorygpustack/gpustack
Ownergpustack
Primary languagePython
LicenseApache-2.0 — OSI-approved
Stars5.3k
Forks565
Open issues614
Latest releasev2.2.1 (2026-07-03)
Last updated2026-07-08
Sourcehttps://github.com/gpustack/gpustack

What gpustack is

Python-based cluster orchestrator supporting NVIDIA, AMD, Ascend, and other accelerators. Manages inference engine lifecycle, handles GPU resource allocation, provides OpenAI-compatible APIs, integrates Prometheus/Grafana for observability, and supports speculative decoding and extended KV cache optimizations. Architecture separates control plane (CPU-only) from distributed worker nodes.

Quickstart

Get the gpustack source

Clone the repository and explore it locally.

terminalbash
git clone https://github.com/gpustack/gpustack.gitcd gpustack# follow the project's README for install & configuration

Need it deployed, integrated, or customized instead? DEV.co ships production installs.

Best use cases

Multi-Cloud GPU Inference Platform

Organizations needing to serve LLMs and other AI models across heterogeneous GPU clusters (on-prem, Kubernetes, cloud providers) with unified management and automatic engine configuration.

Enterprise Model-as-a-Service (MaaS) Deployment

Service providers requiring authenticated, metered, load-balanced API access to models with built-in monitoring, access control, and token usage tracking via OpenAI-compatible endpoints.

High-Throughput and Low-Latency Inference

Teams optimizing inference performance through pluggable engine selection, speculative decoding (EAGLE3, MTP, N-grams), extended KV cache (LMCache, HiCache), and pre-tuned throughput/latency modes.

Implementation considerations

  • Requires Docker and NVIDIA Container Toolkit (or equivalent for other accelerators) on all worker nodes; Linux mandatory for workers.
  • Server control plane can run on CPU-only node or co-located with a GPU worker; plan network topology for multi-cluster federation.
  • Initial setup includes cluster registration, model catalog selection, and API key provisioning; no additional complex configuration for basic deployments.
  • Supports broad accelerator ecosystem (NVIDIA, AMD, Ascend, Hygon, MThreads, Iluvatar, MetaX, Cambricon, T-Head) but hardware-specific driver/toolkit setup is a prerequisite.
  • Performance tuning available through pre-tuned modes and advanced parameters; baseline performance shown in benchmarks but workload-specific tuning may be required.

When to avoid it — and what to weigh

  • Non-Linux Worker Nodes Required — GPUStack worker nodes only support Linux. Windows/macOS are not supported for workers (server can run on Docker Desktop, but this limits scalability).
  • Single-Node, Single-GPU Scenarios — If you need a lightweight single-machine solution without multi-cluster orchestration, simpler alternatives (e.g., direct vLLM/SGLang) may have lower operational overhead.
  • Proprietary Inference Engine Requirement — GPUStack's pluggable architecture targets standard inference engines. If your workflow requires deeply integrated proprietary inference engines, compatibility must be verified.
  • Minimal Operational Overhead Needed — GPUStack introduces cluster management, authentication, monitoring, and metering complexity. Teams wanting zero infrastructure overhead should avoid this.

License & commercial use

Apache License 2.0 (Apache-2.0). Permissive OSI-approved open-source license permitting commercial use, modification, and distribution with attribution and liability/warranty disclaimers.

Apache-2.0 clearly permits commercial use, modification, and redistribution. No restrictions on building proprietary products on top. However, any modifications to GPUStack itself must retain Apache-2.0 licensing; consult legal review if bundling or modifying core dependencies. No commercial support, SLA, or indemnification claimed from the open-source project.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Authentication and access control mentioned as enterprise-grade features (API key generation, per-user access). Network isolation between control plane and workers relies on secure token exchange. No audit trail, encryption-at-rest, or network encryption details provided in available data. Requires review: RBAC granularity, audit logging, secrets management, supply-chain security of pulled models and engines. Security posture depends on underlying Docker/Kubernetes/Linux hardening and GPU driver updates.

Alternatives to consider

vLLM (standalone)

Direct inference engine without orchestration; simpler for single or small clusters but lacks multi-cluster federation, authentication, monitoring, and managed model lifecycle.

Ray Serve + vLLM/SGLang

Distributed inference framework with custom serving logic; more flexible but requires more DIY operational work for clustering, model selection, and monitoring compared to GPUStack's built-in features.

Kubernetes + KServe + vLLM

Cloud-native orchestration with standard Kubernetes tooling; good for organizations already invested in K8s but steeper learning curve and requires more manual engine/model configuration than GPUStack's opinionated defaults.

Software development agency

Build on gpustack with DEV.co software developers

Evaluate GPUStack for your AI inference infrastructure. Review the documentation, run the quick-start on a test cluster, and assess multi-cluster federation fit for your environments.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

gpustack FAQ

Can GPUStack run on Windows or macOS?
The server can run on Docker Desktop (Windows/macOS), but worker nodes must be Linux only. For Windows development, WSL2 is recommended; macOS is not supported for workers.
What GPU types are supported?
NVIDIA, AMD, Ascend, Hygon DCU, MThreads, Iluvatar, MetaX, Cambricon MLU, and T-Head PPU. Hardware-specific driver and container toolkit setup is required; see installation requirements documentation.
Is there commercial support or an SLA?
No. GPUStack is open-source under Apache-2.0 with community support (Discord, GitHub issues). Commercial support is not claimed; users must self-support or engage third-party providers.
Can I use proprietary models with GPUStack?
Yes, provided the model is compatible with the inference engines (vLLM, SGLang, TensorRT-LLM, etc.) GPUStack orchestrates. You control which models are deployed; catalog is optional.

Software development & web development with DEV.co

Need help beyond evaluating gpustack? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and ai frameworks integrations — and maintain them long-term.

Ready to Orchestrate Your GPU Clusters?

Evaluate GPUStack for your AI inference infrastructure. Review the documentation, run the quick-start on a test cluster, and assess multi-cluster federation fit for your environments.