v2.9.0

May 9, 2026

Major features

Add HAMi-core mode for Ascend devices, enabling user-space virtualization for fine-grained memory and compute sharing.
Optimize HAMi-core performance and add the latest benchmark data for HAMi-core.
HAMi-DRA for NVIDIA is ready for use.
Sync Volcano vGPU Device Plugin with version 0.19 and add CDI support.
Add HAMi skills for debugging and development workflows.
Support module-pair allocation for Ascend 910C devices in SuperPod environments by (@ashergaga) in #1610
Add support for Vast.ai devices by (@DSFans2014) in #1645
Add Ascend ResourceCoreName and Ascendxxx-core resources to support hami-vnpu-core virtualization by (@ashergaga) and (@DSFans2014) in #1771 and #1804
Support node filtering based on hami-vnpu-core annotations and multi-device requests with hami-vnpu-core enabled by (@ashergaga) in #1812 and #1837

Fix initialization errors when using tensor parallelism on vLLM versions greater than 0.18.
Fix schedulerName precedence checks by (@hoteye) in #1627
Add nil checks to prevent leader election panics by (@haitwang-cloud) in #1603
Fix panic on nil resource requests in scheduler scoring by (@yxxhero) in #1626
Fix reversed binpack and spread scheduling policies for Iluvatar devices by (@qiangwei1983) in #1631
Resolve cardinality explosion in Device_memory_desc_of_container by (@maishivamhoo123) in #1628
Handle GetMemoryInfo ERROR_NOT_SUPPORTED for unified memory GPUs by (@jsl9208) in #1637
Optimize nodelock scalability with exponential backoff and listers by (@maishivamhoo123) in #1663
Fix readiness probes when replicas are greater than one by (@Shouren) in #1677
Fix scheduler slot usage prediction and device type filtering by (@maishivamhoo123) in #1700
Retain terminating pods in cache to prevent premature eviction by (@maishivamhoo123) in #1719
Fix multi-container allocation when init containers are present by (@haitwang-cloud) in #1650
Align kubelet allocation with scheduler annotations by (@xrwang8) in #1743
Handle Linux kernel 6.17 handshake edge cases in NVIDIA health checks by (@maishivamhoo123) in #1810
Fix MIG allocation failures in CDI mode by (@DSFans2014) in #1826

Add vGPUmonitor --metrics-bind-address flag by (@dongjiang1989) in #1613
Add Prometheus ServiceMonitor support in Helm charts and device plugins by (@dongjiang1989) in #1614 and #1633
Check resource quota in webhook by (@DSFans2014) in #1605
Add namespaceSelector and objectSelector configuration for the webhook Helm chart by (@haitwang-cloud) in #1653
Align Prometheus metric and label names with best practices by (@MyoungHaSong) in #1644
Optimize log verbosity and add unit tests by (@haitwang-cloud) in #1710
Add local-deploy target for minikube and kind clusters by (@anandj91) in #1760
Add hami_vgpu_metrics_summarizer and k8s-debug-gpu-pod skills by (@haitwang-cloud) in #1755 and #1654
Add DeepCopy functions for DeviceUsage and nested types by (@Shouren) in #1818
Add enableGetPreferredAllocation flag by (@DSFans2014) in #1824
Add device type labels to metrics by (@xiyichan) in #1612
Add io.LimitReader to scheduler routes to prevent denial-of-service risks by (@maishivamhoo123) in #1620
Remove deprecated scheduler policy ConfigMap by (@haitwang-cloud) in #1651
Update NVIDIA device plugin and NVIDIA container runtime modules by (@archlitchi) in #1731
Upgrade Go to 1.26.2 and address related security issues by (@luohua13) and (@Shouren) in #1791 and #1772
Disable host network for the device plugin by (@luohua13) in #1789
Bump HAMi-DRA version to v0.2.0 by (@FouoF) in #1845

Full Changelog: https://github.com/Project-HAMi/HAMi/compare/v2.8.0...v2.9.0