dsx-ai-factory/run-ai

Run:ai is the AI workload orchestration platform for Kubernetes. This monorepo contains the control plane services, in-cluster Kubernetes operator/controllers, shared libraries and clients, the installer, and the local dev environment

Summary

NVIDIA Run:ai is an open-source AI workload and GPU orchestration platform that runs on Kubernetes. This monorepo contains the control plane services, in-cluster controllers, installer CLI, MCP server, shared client packages, and a local development environment. It targets teams that need to share GPU resources across multiple Kubernetes clusters with quotas, policies, and central visibility.

What it is useful for

Run:ai extends Kubernetes scheduling with GPU-aware capabilities such as fair-share allocation, preemption, fractional GPU sharing, and topology-aware placement. It treats AI workloads as first-class types: workspaces, training, distributed training, and inference, and extends support for additional Kubernetes workloads through integrations with Training Operator, LeaderWorkerSet, Knative, and KServe.

The platform fits multi-tenant scenarios where departments or projects share one GPU estate across on-premises, cloud, or hybrid Kubernetes. Administrators can define resource quotas and policies per department/project, while users submit work through the web UI, CLI, or REST API. A single control plane manages workloads, policies, and configuration across connected clusters, and an MCP server makes those operations available as tools for AI assistants.

The README also clarifies where Run:ai does not fit: it runs on Kubernetes and OpenShift rather than replacing them, sits alongside GPU enablement tooling like the NVIDIA GPU Operator, and is complementary to ML platforms such as Kubeflow, Ray, and Argo Workflows. For a single team on dedicated GPUs, native Kubernetes scheduling may be enough.

How engineers can use it

For local exploration, the repo ships a one-command Kind-based environment:

./local-dev.sh
./local-dev.sh --skip-build --skip-cluster
./local-dev.sh --teardown

Prerequisites documented in local-dev.md include Docker or Podman, make, git, curl, python3, and openssl; kind, kubectl, and helm are installed automatically. Go, Node, and Java are not needed on the host because builds run in containers.

For production, the README points to a guided installation using the runai-installer CLI with a configuration YAML, plus a manual Helm chart path suitable for GitOps tools like Argo CD. Released images and Helm charts come from nvcr.io/nvidia/runai by default.

Each service directory has its own Makefile with make build, make test, make lint, and make validate targets. The README explicitly notes that root-level Makefile targets are placeholders, so engineers should work at the service level and consult docs/developer/services.md for the breakdown.

License: Apache-2.0, covering code only. The Apache 2.0 license does not grant rights to NVIDIA or Run:ai trademarks. Vendored third-party code under other licenses is listed in per-subdirectory NOTICE files.