NVIDIA released the full Run:ai platform as open-source software on Oct. 13, giving operators code for scheduling AI workloads and sharing GPUs across Kubernetes clusters. Teams can use the platform as a whole or adopt four components separately for job placement, GPU sharing, team-level resource management and workload descriptions. The announcement reports no before-and-after GPU-performance test, so it establishes software access, not a measured utilization gain. In a same-day follow-up, Lambda said it plans to package Run:ai as a managed option on Lambda Kubernetes Service, but it gave no availability date.
Run:ai adds GPU-aware policy to Kubernetes
Kubernetes is software for deploying and managing applications across machines. Run:ai extends it with GPU-aware scheduling and resource policies rather than replacing it. The code is aimed at shared AI clusters where many teams’ training, inference and development jobs compete for accelerators.
That layer can decide which job runs where, how teams share capacity and how multi-GPU jobs are placed together. It is most relevant when an organization needs quotas, fair access or fractional GPU allocation across teams; a single team with dedicated GPUs may not need another scheduling layer.
Four components cover scheduling, GPU sharing and workload descriptions
- KAI Scheduler places jobs using policies such as fair share and topology awareness. For a multi-pod job, gang scheduling tries to place the whole group together or not at all, rather than starting only part of it.
- GPU Fractioning Operator connects CUDA 13.4’s fractional GPU allocation to Kubernetes. NVIDIA says the CUDA driver enforces a hard memory limit for each workload’s assigned fraction.
- KAI Resource Management organizes capacity across projects, departments and node pools, with quotas, guarantees and shared allocations.
- Workload Map records a job’s pods, components and roles in a shared format. That gives the scheduler and other platform tools a common way to discover and track different job types without building a separate integration for each. Workload Map describes workload structure; it does not schedule, queue or autoscale jobs.
NVIDIA says the components can be used independently or with additional platform code as an integrated Run:ai deployment. That modularity lets a team fill a specific gap without adopting every part, while an operator can combine scheduling with organizational quotas for a multi-team cluster.
Lambda’s managed-service integration is still a plan
Lambda says it is working to make Run:ai a first-class option in its managed Kubernetes service. Under the proposed split, Run:ai would schedule AI workloads and allocate GPU capacity, while Lambda’s service would manage Kubernetes and monitor or remediate unhealthy GPU nodes. Lambda says it has supported customer-deployed Run:ai on its infrastructure since 2022, but its new integration announcement describes future work, not an already available managed option.
Open-source access has separate license and support terms
The source is on GitHub, where the repository lists Apache 2.0 at its root and separate license notices for some nested components. The NGC catalog also offers an integrated Helm-chart package; it says an account, subscription or API key is not required to download the artifacts. NVIDIA’s launch post says the NGC Community Product remains subject to applicable NVIDIA software and product terms, is community-supported and does not include NVIDIA enterprise support. Commercial NVIDIA offerings continue under their own terms.
Image credit: Photograph by Carl Lender via Wikimedia Commons, used under CC BY 2.0. It is illustrative and does not show a Run:ai deployment. Image source and license.
.jpg){kind=link}