Heterogeneous Management
Manage and schedule GPU, NPU, MLU, and other accelerators in one workflow.
CNCF: IncubatingLatest Release: v2.10.0
Heterogeneous GPU Sharing on Kubernetes
HAMi is an open-source, cloud-native GPU virtualization middleware that brings sharing, isolation and scheduling of heterogeneous accelerators to AI workloads on Kubernetes.
GPU SlicingHeterogeneous AcceleratorsKubernetes-native Scheduling
AI Workloads
LLM
ML
HPC
Kubernetes Scheduling Ecosystem
Virtualization • Sharing • Isolation • Scheduling
GPU
1/2
1/4
1/N
Heterogeneous Accelerators
and more...
CNCF Incubating Project
HAMi is a CNCF Incubating project
HAMi is an Incubating project of the Cloud Native Computing Foundation (CNCF), listed in both the CNCF Landscape and the CNAI Landscape.
Why HAMi
Hard Isolation
Slice memory and compute precisely with hard isolation at runtime.
Advanced Scheduling
Use binpack, spread, and topology-aware policies for better placement.
Kubernetes Native
Work with Kubernetes APIs, DRA, and CDI for easier adoption.
Resource Isolation & QoS
Control memory and core quotas for fairer and more stable sharing.
Unified Monitoring
Provide consistent metrics and visibility across device vendors.
Architecture & How It Works
HAMi runs in Device Plugin or DRA mode. Both are used the same way. The difference is who schedules: HAMi, or Kubernetes.
Device Pluginclassic HAMi
DRAHAMi-DRA
HAMidevice pluginnode annotations
HAMiDRA driverResourceSlice
HAMi webhooksets schedulerName
HAMi-DRA webhookconverts to ResourceClaim
hami-schedulerkube-scheduler + extender
Kuberneteskube-schedulercluster default
Pod annotationsplus a node lock
ResourceClaim status.allocation
device pluginAllocate, then find the Pod
HAMiDRA driverthrough CDI
HAMi-corein the container
HAMi-corein the container
Before and After Using HAMi
Compare traditional whole-GPU allocation with HAMi GPU sharing under the same workloads.
Ecosystem & Device Support
Broad accelerator ecosystem across vendors. See docs for full support matrix.
Works with HAMi
HAMi integrates with these open-source schedulers, queuing layers, and cloud platforms, either as an embeddable device layer or at the scheduling layer.
Adopters
The organizations below are evaluating or using HAMi in production environments.
Join the adopters list
Submit your organization through the contributor guide process.
See submission instructions →Technical Partners
These organizations provide us with hardware, resources, and top tier support. This is a deep technical partnership between the HAMi maintainers and these organizations. If you are interested in this level of commitment, please contact us.
Contributors
HAMi is advanced by contributors from the community and industry. These organizations actively participate in project development and ecosystem collaboration.
Global Community Metrics
A live snapshot of HAMi community growth and open-source momentum.
