dstack
Cost-effective LLM development
FreeOpen SourceAIFramework6,660 impressions#14 of its week
dstack is an open-source orchestration layer for AI workloads across clouds, Kubernetes, VMs, and bare-metal.
- for
- AI researchers, data-center operators, and teams needing multi-cloud GPU orchestration.
- pricing
- open source
- license
- MPL-2.0
Key features
- Unified control plane — Manage compute, training, and inference across GPU clouds, Kubernetes, Slurm, and bare-metal from one interface.
- Fleet provisioning — Create and monitor SSH fleets, Kubernetes clusters, or Slurm backends for heterogeneous accelerators.
- Task scheduling — Schedule training, inference, and custom jobs with first-class primitives for AI workloads.
- Service layer — Deploy cache-aware, PD-disaggregated inference services with auto-scaling and rate limits.
- Presets toolkit — Agent-based optimization presets to streamline resource usage and cost.
- Tenant isolation — Projects provide usage metering and isolation for multi-tenant AI labs or token factories.
Use cases
- Run multi-node training experiments on multiple GPU clouds without vendor lock-in.
- Deploy inference services on existing on-prem Kubernetes or Slurm clusters.
- Provide a shared AI compute platform for internal teams across cloud and bare-metal resources.
dstack FAQ
Can I use dstack with Kubernetes?+
Yes, connect existing Kubernetes clusters through the Kubernetes backend and dstack will schedule AI workloads alongside other resources.
Can I use dstack with Slurm?+
Yes, the experimental Slurm backend lets dstack submit jobs to Slurm clusters via SSH, while retaining Slurm's scheduling.
What problem does dstack solve compared to Slurm?+
dstack adds a unified, AI-focused orchestration layer with primitives for compute, training, inference, and observability, usable on cloud GPUs, VMs, or bare-metal.
What problem does dstack solve compared to Kubernetes?+
It provides a lightweight, AI-specific interface on top of container orchestration, simplifying compute management for heterogeneous AI workloads.
How does dstack reduce cloud costs?+
By unifying control of multiple GPU clouds and on-prem resources, it enables efficient utilization and avoids over-provisioning.
Where can I get help with dstack?+
Support is available via the project's Discord channel or direct contact links on the website.
Summarized by DevHunt from dstack.ai · Sep 30, 2026. Details may change; check the official site.
.gif1688985694192.png?width=64)
