# hami v2.7.1

2026-01-23 · [Original release notes](https://github.com/Project-HAMi/HAMi/releases/tag/v2.7.1) · [Full summary](https://ratatosk.io/en/releases/hami/v2.7.1)

Security 1 · Breaking changes 2 · Important fixes 8 · Dependency updates 1 · Other changes 12

## Security (1)

### `nvidia-mig-parted` security fix in Docker builds

- Affects: images built from `docker/Dockerfile`, `docker/Dockerfile.hamimaster`, or `docker/Dockerfile.withlib` (Inferred)
- [CVE-2025-23266](https://www.cve.org/CVERecord?id=CVE-2025-23266)

These builds previously installed `nvidia-mig-parted` v0.10.0, which was affected by the vulnerability. They now install the fixed version, v0.12.2.

> * fix: upgrade nvidia-mig-parted to v0.12.2 to solve security issues by @Shouren in https://github.com/Project-HAMi/HAMi/pull/1388

## Breaking changes (2)

### Cambricon resource-assigned key spelling corrected

- Affects: HAMi setups using the Cambricon device implementation (Inferred)

The code and test previously used `CAMBRICON_DSMLU_ASSIGHED`; both now use `CAMBRICON_DSMLU_ASSIGNED`.

> * fix: fix typos by @DSFans2014 in https://github.com/Project-HAMi/HAMi/pull/1434

### Iluvatar chart support now defaults to disabled

- Affects: HAMi installations using or configuring Iluvatar GPUs

The chart previously exposed generic Iluvatar resource names; it now offers `devices.iluvatar.customresources` for model-specific resources. The old `iluvatar-name`, `iluvatar-memory` and `iluvatar-cores` flags are replaced by `enable-iluvatar`.

Evidence: “This PR improves support for Iluvatar GPUs, including device plugin (gpu-manager) adaptation, scheduling policy compatibility, and resource isolation and sharing capabilities.” https://github.com/Project-HAMi/HAMi/pull/1399

> * Improved support for iluvatar GPUs by @qiangwei1983 in https://github.com/Project-HAMi/HAMi/pull/1399

## Important fixes (8)

### Terminated-pod device quota cleanup fixed

- Affects: pods whose devices have tracked quota usage

When replicas changed from 1 to 0, the scheduler could process the update event before the deletion event and delete the pod from its manager first, leaving its device quota usage uncleared. The scheduler now removes quota usage from the pod’s recorded devices before deleting it from the pod manager.

> **Fix: Calculation error for quotas by @luohua13 in https://github.com/Project-HAMi/HAMi/pull/1400**

### GPU plugin-removal health check fixed

- Affects: GPU nodes where the scheduler checks the `hami.io/node-handshake` annotation

When the device plugin was removed from a GPU node, `CheckHealth` returned `true, true` because the scheduler and plugin updated different prefix values for the annotation. The health check now reads the annotation through `util.HandshakeAnnos[devType]`.

> * Fix: After removing the device plugin from the gpu node, it can still… by @luohua13 in https://github.com/Project-HAMi/HAMi/pull/1456

### HAMi-core updated to address vLLM-related issues

- Affects: vLLM 0.9.2 on HAMi vGPU with only virtual GPU index 1 exposed (`CUDA_VISIBLE_DEVICES=1`)

In this setup, vLLM failed during memory profiling and aborted startup with a `host pid is error!` message.

> **Update HAMi-core to fix vllm-related issues: #1381 # 1461 by @archlitchi in https://github.com/Project-HAMi/HAMi/pull/1478**

### `Resourceqota` hard-quota clearing fixed

- Affects: `Resourceqota` custom resources

When the hard quota was cleared while the custom resource remained, the update handler used the new object to delete quota data, leaving the existing quota uncleared. The scheduler now deletes quota data using the old object before adding it using the updated object.

> **Fix: Calculation error for quotas by @luohua13 in https://github.com/Project-HAMi/HAMi/pull/1400**

### Helm ClusterRoleBinding role reference name fixed

- Affects: Helm chart installations with a release name or chart name other than `hami`

In these installations, the static `roleRef.name` did not match the ClusterRole name, causing the ClusterRoleBinding to fail.

> * Fix: failed clusterrolebinding when change release name or chart name by @FouoF in https://github.com/Project-HAMi/HAMi/pull/1380

### Nil `pod` check added to `ReleaseNodeLock`

- Affects: code that calls `ReleaseNodeLock` (Inferred)

In some cases, passing a nil `pod` to `ReleaseNodeLock` could cause a panic.

> * fix: check pod nil in `ReleaseNodeLock` by @DSFans2014 in https://github.com/Project-HAMi/HAMi/pull/1372

### `GetScheduledPods()` returns a shallow copy of the pod map

- Affects: HAMi scheduler (Inferred)

Concurrent iteration and writes to the scheduler's pod map caused a fatal error.

> * Fix concurrent map iteration and map write fatal error. by @litaixun in https://github.com/Project-HAMi/HAMi/pull/1452

### Concurrent map access fix in `nodeManager.ListNodes()`

- Affects: HAMi scheduler code that calls `nodeManager.ListNodes()` (Inferred)

Concurrent reads and writes to the node map could cause a fatal error. Previously, `ListNodes()` returned the internal map; it now returns copied node information and omits entries with a nil `NodeInfo` or nil `Node`.

> * Fix concurrent map read write fatal error. by @litaixun in https://github.com/Project-HAMi/HAMi/pull/1476

## Dependency updates (1)

1 more, none mentioning security

- `* Fix release CI by @archlitchi in https://github.com/Project-HAMi/HAMi/pull/1373`

## Other changes (12)

- Commented-out HAMI image-build step removed from the release workflow
- E2E test script uses the Ginkgo version from `go.mod`
- Scheduler resource-quota test waits for pod deletion to be processed
- Spelling corrected in HAMi diagnostics and metric description
- Spelling corrections across project documentation, examples, tests, and comments
- NVIDIA helper names corrected
- CI and NVIDIA MIG test updates
- `DeviceUsage` now includes pod information
- NVIDIA GPU NUMA-node acquisition switched to sysfs
- Node and Pod annotation patches use `MergePatchType`
- Vendor custom scheduling-failure reasons supported
- HAMi Helm chart default image tag changed to `v2.7.1`

---

Summaries are produced by AI, so they can be wrong, and they are provided without warranty.
