Back to Gap Map Infrastructure / ML orchestration

KAI Scheduler

open source / Overall score: 3.2

Kubernetes-native scheduler for AI workloads at large scale. It schedules whole PodGroups rather than individual pods, supports hierarchical queues with fair-share over time, bin-packing and spread placement, topology-aware scheduling, and separates workload priority from preemptibility. It runs alongside other schedulers on the same cluster rather than replacing them.

Originated at Run:ai and was published by NVIDIA; the repository now lives in its own kai-scheduler organization rather than under NVIDIA, and the github.com/NVIDIA/KAI-Scheduler path redirects there. Scored as its own product, separately from the closed NVIDIA Run:ai platform.

Openness

5 high confidence

5.0

license: Apache-2.0(OSI)
source: public(the scheduler you run builds from this repo)
core features withheld: no — a 2,004-entry recursive tree carries one root LICENSE and no path matching enterprise, ee, commercial or proprietary

KAI Scheduler publishes the whole scheduler under Apache-2.0. NVIDIA sells Run:ai beside it as a separate product rather than a fuller edition of this one, and a full scan of the repository tree finds no enterprise, ee or commercial path and no license-key machinery.

https://api.github.com/repos/kai-scheduler/KAI-Scheduler/git/trees/main?recursive=1 recorded 2026-09-15
Full untruncated recursive tree of the default branch, 2,004 entries. One root LICENSE. No path matches enterprise, ee, commercial or proprietary. A tree lists paths; it is cited for that and for nothing about the vendor's offerings.
https://github.com/kai-scheduler/KAI-Scheduler recorded 2026-09-15
Repository page for kai-scheduler/KAI-Scheduler: Apache-2.0 license, public and unarchived, in the kai-scheduler organization rather than NVIDIA's. Establishes the license, the source, and the current repository identity.
https://www.nvidia.com/en-us/software/run-ai/ recorded 2026-09-16
NVIDIA's Run:ai page, which describes the open KAI Scheduler as 'based on NVIDIA Run:ai' and sells Run:ai as a separate enterprise platform. Establishes that the commercial product is a different thing rather than a fuller edition of this scheduler.

Adoption

2 low confidence

2.0

KAI Scheduler ships as container images and a Helm chart rather than through a package registry, leaving GitHub stars as the only adoption signal. A star is not a use, so the signal shows attention rather than real usage.

https://github.com/kai-scheduler/KAI-Scheduler recorded 2026-09-15
Repository page for kai-scheduler/KAI-Scheduler showing 1,520 stars.

Capability

5 medium confidence

5.0

KAI Scheduler decides which nodes a PodGroup lands on and admits the whole group as one allocation, holding the machines it reserves together rather than deferring that choice elsewhere. Volcano performs the same function on the same substrate.

https://github.com/kai-scheduler/KAI-Scheduler recorded 2026-09-15
Repository page and README describing KAI Scheduler as a Kubernetes-native scheduler for AI workloads, with hierarchical PodGroups, bin-packing and spread scheduling, topology-aware scheduling and time-based fair-share.

Verified 2026-09-15