Nvidia has released its complete GPU orchestration platform as open source under Apache 2.0, publishing the code on GitHub through its DSX AI Factory organisation. The move hands production scheduling and quota management to anyone willing to run it.
The repository is a monorepo holding a control plane, cluster services, an installer, a command line tool and an MCP server for AI assistants. A local build compiles roughly 50 Go services from source.
Four pieces teams can adopt separately
KAI Scheduler is the Kubernetes-native engine, a CNCF Sandbox project covering gang scheduling, fair-share allocation and topology awareness. A GPU Fractioning Operator brings CUDA 13.4 native fractions with driver-enforced memory limits, replacing Nvidia’s older proprietary approach.
KAI Resource Management adds projects, departments and quotas, with worked examples guaranteeing half a cluster to one research team. Workload Map describes job types like PyTorchJob and RayCluster but does not schedule anything itself.
Development does not require a GPU. A single script builds into a five-node Kind cluster pinned to Kubernetes v1.32.5 and simulates eight Tesla-A10 cards, so a laptop on macOS or Linux can run the whole stack. Windows is not supported.
The community edition carries no Nvidia enterprise support, and AMD and Intel accelerators are out of scope. Nvidia also pointed to Model Streamer, which cut model load times three to six times on eight H100 virtual machines in a Microsoft benchmark.
The comparison that matters is scope. Volcano, Kueue and YuniKorn are all Apache 2.0 but focus on scheduling or queueing alone. This stack ships the management layer above the scheduler, which is the part enterprises usually pay for.
