NVIDIA's announcement and public repository confirm the full platform code, component scope, Apache 2.0 license and community-support boundary.
NVIDIA extends Run:ai with GPU controls and Workload Map reuse
2026-10-13Verified announcementTRIAL
NVIDIA has combined memory-enforced GPU fractioning in Run:ai with a reusable Workload Map layer for heterogeneous Kubernetes AI workloads. Workload Map converts supported CRDs into a common tree used by Run:ai, KAI Scheduler and, according to NVIDIA, Elotl Nova. The library is Apache-2.0 and released, but production-use and interoperability evidence remains vendor-reported.
Official · ReferenceNVIDIA Run:ai is now open sourceOfficial · ReferenceRun:ai open-source platform repository
Availability: public GitHub monorepo and NGC community distribution
What this means for you
For platform teams, a shared workload translation layer can reduce one-off integrations across schedulers, inspection and multi-cluster placement. Trial it in isolation: custom CRDs, replica/resource accounting and status propagation still need independent validation, while GPU-sharing failure boundaries remain separate concerns.
Evidence & related material
Run:ai adds CUDA 13.4 memory-enforced GPU fractioning.
NVIDIA documents the Workload Map architecture and reports that Run:ai, KAI Scheduler and Elotl Nova reuse its common workload tree.
Subsequent changes
Run:ai adds CUDA 13.4 memory-enforced GPU fractioning
NVIDIA documents Workload Map architecture and three reuse paths
Technical details & scores
| Availability | public GitHub monorepo and NGC community distribution |
|---|---|
| Components | control plane, cluster services, installer, CLI, MCP server, KAI Scheduler, GPU Fractioning Operator, KAI Resource Management, Workload Map |
| License | Apache-2.0 for the GitHub repository; NGC use remains subject to NVIDIA terms |
| Support | community build has no enterprise support entitlement |
| Time Basis | official blog published 2026-10-13T14:19:40Z; discovered inside the configured overlap and not present in the previous manifest |
| Gpu Fractioning Update | CUDA 13.4 memory-enforced fractioning with request, limit and burst allocation; time-slicing uses separate error domains while SM sharing uses a common error domain |
| Published At | 2026-10-27T20:50:39Z |
| Workload Map Update | Official architecture disclosure published 2026-11-02T16:00:00Z. Apache-2.0 CRD and Go library map JobSet, PyTorchJob, Ray, KServe, Knative, MPIJob, NIM, LeaderWorkerSet, Milvus, Grove and Dynamo into a common workload tree. NVIDIA reports production Run:ai use, KAI Scheduler fallback use and Elotl Nova integration; v0.3.0 itself was released 2026-10-21. |
Technical impact 4
Engineering value 4
Adoption signal 3
Maturity 3
Verification cost 4
Risk 3