AI Potluck
Contribute
Back to Gap Map Infrastructure / ML orchestration

HAMi

Project HAMi
open source / Overall score: 2.8

Heterogeneous GPU sharing and scheduling for Kubernetes, hosted by the CNCF. HAMi registers each accelerator in a cluster with its split count, memory limit and core limit, and its own scheduler decides which node and which devices a pod receives, recording the decision on the pod. It covers NVIDIA GPUs alongside Cambricon MLUs and other vendors' accelerators.

Ships as container images and a Helm chart with no package channel, so adoption bands on stars. Verified 2026-09-16 via the repository, its recursive tree, the HAMi scheduling protocol and the CNCF project page.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI)
source
public(Project-HAMi/HAMi is the product you run)
core-gated
ungated(a 591-entry recursive tree carries no path matching enterprise, ee, commercial or proprietary)

Apache-2.0 under CNCF hosting, with no vendor positioned to sell a fuller edition. The recursive tree carries no enterprise, ee or commercial path.

  • https://github.com/Project-HAMi/HAMi recorded 2026-09-16

    Repository page for Project-HAMi/HAMi: Apache-2.0 license, public and unarchived, described as 'Heterogeneous GPU Sharing on Kubernetes'. Establishes the license and that the source is public.

  • https://api.github.com/repos/Project-HAMi/HAMi/git/trees/master?recursive=1 recorded 2026-09-16

    Full untruncated recursive tree of the default branch, 591 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://www.cncf.io/projects/hami/ recorded 2026-09-16

    HAMi's project page on cncf.io, establishing CNCF hosting rather than vendor ownership, so no commercial party is positioned to withhold functionality from the published source.

Adoption

2 low confidence
2.0

4,600 GitHub stars. HAMi ships as container images and a Helm chart and publishes no package-registry artifact, so no usage-volume channel exists and the band rests on stars, which measure attention rather than use. The stars route caps at level 3.

Capability

4 medium confidence
4.0

Admitted on the accelerator limb: HAMi registers devices in its own annotation schema and its own scheduler consumes that to choose placement. Band 4 - it selects the node AND the specific devices, which is more than dispatching, and nothing cited shows it reserving a multi-machine set as one allocation, which is what band 5 asks. Repointed from nvidia-run-ai to openpai when Run:ai moved to band 5: a peer must be at the band being claimed, and openpai selects hardware on the same footing this record does.

Verified 2026-09-16