#Release Notes

Use Versions and Components for the current marketplace versions, component roles, and installation links. The sections below record the changes and limitations of each release.

For NVIDIA GPU metrics changes, see the Alauda Build of DCGM-Exporter release notes in the NVIDIA GPU documentation. For Ascend NPU metrics changes, see the NPU Exporter information in the Ascend NPU documentation.

#Alauda Build of HAMi

#v2.9.0

Alauda Build of HAMi v2.9.0 is the current formal Alauda release for HAMi sharing, virtualization, and scheduling. It is delivered as an ACP cluster plugin with amd64, arm64, and multi-architecture (ALL) packages.

#New features and improvements

• Added the complete custom kube-scheduler image path for deployments that need to select a scheduler image explicitly.
• The Alauda delivery package configures standard tolerations for HAMi scheduler and device-plugin workloads on tainted accelerator nodes.
• Refreshed the formal HAMi chart, runtime image, and ACP plugin packages from the v2.9.0 release tag.

#Upgrade notes

• Before upgrading, confirm the target ACP version, backend path, node taints, and currently exposed HAMi resource keys.
• If a custom scheduler image is required, configure scheduler.kubeScheduler.customImage through the HAMi plugin parameters; when it is empty, the image is generated from the target cluster Kubernetes version and configured registry.
• After upgrading, verify HAMi scheduler and device-plugin readiness, node allocatable resource keys, ACP quota metadata, and one representative workload from Verification.

#v2.8.3

#Fixed issues

• Fixed memory information handling for unified memory GPUs when GetMemoryInfo returns NOT_SUPPORTED.
• Fixed HAMi-Core monitoring issues in v2.8.2.
• Fixed device utilization watcher startup when a workload does not specify nvidia.com/gpucores.

#v2.8.1

#New features and improvements

• Added Huawei Ascend NPU sharing support.
• Added CDI mode support for NVIDIA devices.
• Synchronized with NVIDIA Kubernetes Device Plugin v0.18.0.
• Added hami_build_info Prometheus metrics and version output.
• Added TLS certificate watch and hot reload without restarting pods.
• Added NVIDIA GPU Operator toolkit readiness checks.
• Added GPUDirect RDMA copy (GDRCopy) and GPUDirect Storage (GDS) configuration support.
• Added mock device plugin support for testing environments.
• Upgraded HAMi-WebUI to v1.10.0 for HAMi v2.8 compatibility. HAMi-WebUI v1.10.0 is compatible with HAMi v2.7 and v2.8. HAMi-WebUI v1.5.0 is not compatible with HAMi v2.8.

#Fixed issues

• Updated HAMi-Core to fix vLLM-related issues.
• Fixed launch failures for vLLM versions later than 0.18 when multiple GPUs are used.
• Fixed quota calculation errors.
• Fixed MIG instance allocation errors where the scheduler selected incorrect MIG instances.
• Upgraded nvidia-mig-parted to v0.12.2 for security fixes.
• Fixed stale GPU node visibility after the device plugin is removed from a GPU node.
• Fixed concurrent map read/write errors.
• Fixed device NUMA acquisition logic.
• Fixed ClusterRoleBinding errors when changing the release name or chart name.

#v2.7.1

#New features and improvements

• Added NVIDIA GPU ResourceQuota support for ACP 4.2 and later.
• Added aggregated scheduling failure events.
• Made node lock timeout configurable.

#Fixed issues

• Fixed an issue where a GPU node could still be scheduled after the device plugin was removed from that node.

#v2.6.1

#Fixed issues

• Fixed device memory accounting when workloads allocate memory with cuMallocAsync.
• Fixed device memory accounting for gpu_burn.
• Fixed a segmentation fault in some scenarios.
• Fixed utilization metrics when workloads use multiple devices.
• Fixed an initialization error when using vLLM with tensor parallelism greater than 2.

#v2.6.0

#New features and improvements

• Optimized scheduler logs.
• Added support for Enflame GCU sharing.
• Added support for Metax GPU and Metax sGPU.
• Added Helm chart checksum annotations so HAMi components restart after ConfigMap changes.
• Added support for RuntimeClass with NVIDIA devices.
• Added profiling through net/http/pprof.
• Added NVIDIA GPU topology score registration on nodes.
• Added MIG information metrics in vGPU monitoring.

#Fixed issues

• Fixed compatibility with NVIDIA driver 570+.
• Fixed device memory accounting in ComfyUI tasks.
• Fixed Cambricon device allocation.
• Fixed incorrect logs and container requested-device count errors.
• Fixed inconsistent vgpu-devices-allocated annotations.
• Fixed stale node devices in node manager.
• Fixed missing single-GPU granularity in dynamic GPU partitioning.
• Fixed device memory accounting on cuMallocAsync.
• Fixed scheduler crash when a MIG task accidentally runs on a hami-core GPU.
• Fixed multi-process device memory accounting.

#Alauda Build of HAMi-WebUI

#v1.10.3

Release date: 2026-07-27

Alauda Build of HAMi-WebUI v1.10.3 is a patch release for HAMi v2.9.x. It is available from Customer Portal and AC CN/IO as amd64, arm64, and multi-architecture (ALL) cluster-plugin packages. The plugin uses the upstream HAMi-WebUI v1.2.0 application baseline.

#Fixed issues

• Changed task-list name filtering to search the Pod or workload name represented by each row. An internal container name is not a task-list name filter.
• Fixed device-allocation annotation decoding for Pods that have init containers. Allocation positions now follow the complete init-container plus regular-container order.
• Corrected MLU device-allocation annotation decoding.

#Security

• Upgraded the Go toolchain, gRPC, and golang.org/x/* dependencies to clear all currently fixable backend findings.
• Upgraded vulnerable frontend dependencies, including body-parser, without changing the plugin form or WebUI workflow.
• The released backend and frontend images passed the Critical and High vulnerability gates. CVE-2026-6993 in go-kratos/kratos v2.7.3 remains a Medium finding because no fixed upstream release is available.

#Upgrade notes

• Upgrade from v1.10.2 through the ACP cluster-plugin page. Preserve the Prometheus address and NodePort choice.
• Regenerate the Prometheus authentication value from the current monitoring Secret. The value entered in the plugin form must include the Basic prefix.
• After upgrade, verify that the backend Prometheus query returns JSON and that a Pod-name task filter returns the expected row.

#Known limitations

• HAMi-WebUI v1.10.3 queries the HAMi 2.9 metric names. Do not use it with HAMi 2.8.x; otherwise affected utilization views can remain at zero.
• The task list is Pod-level. The name filter intentionally does not search internal container names.

#v1.10.2

Alauda Build of HAMi-WebUI v1.10.2 introduced the HAMi 2.9 metric-name update and is superseded by v1.10.3 for new HAMi 2.9.x deployments.

#New features and improvements

• Added cascade filter dropdowns so filter options are scoped to the current selection.
• Optimized monitor overview charts and added workload insights.
• Improved node detail page chart display and fixed time picker internationalization.
• Updated metrics for HAMi v2.9.0 compatibility.

#Fixed issues

• Made /metrics scrape-safe and reduced page-side query fanout to improve large-cluster performance and avoid metrics timeout or out-of-memory issues.
• Fixed blank pages caused by a broken lodash dependency release by restoring lodash 4.17.21.

#Security

• Upgraded the backend to Go 1.26.3 to address 8 Go standard library CVEs: CVE-2026-33811, CVE-2026-33814, CVE-2026-39820, CVE-2026-39823, CVE-2026-39825, CVE-2026-39826, CVE-2026-39836, and CVE-2026-42499.
• Upgraded frontend dependencies, including axios 1.16.0, follow-redirects 1.16.0, postcss 8.5.10, qs 6.15.2, and ws 8.20.1, to address 23 Node CVEs.

#Upgrade notes

• The supported upgrade path is v1.10.0 to v1.10.2. No breaking changes are documented for this release.

#Alauda Build of HAMi Ascend Device Plugin

#v1.4.0

Release date: 2026-07-29

Alauda Build of HAMi Ascend Device Plugin v1.4.0 is a Technical Preview Operator plugin for deploying and managing HAMi Ascend whole-card allocation, hard slicing, and soft slicing.

#New features and improvements

• Added per-node controls to exclude devices from HAMi management and configure the number of virtual devices.
• Automatically enables device sharing on nodes configured for soft slicing.
• Added built-in Prometheus monitoring for soft-slice workloads, with optional ServiceMonitor integration.
• Added vir05_1c_16g and vir10_3c_32g hard-slice templates for Ascend 910C.
• Improved device allocation for multi-container Pods.

#Known limitations

• A hard-slice vNPU can be reclaimed before the workload first accesses the device, even while its Pod is running. Use whole-card mode for workloads that require stable device access.
• A partial-memory hard-slice request supports only one device per container. A request for more than one device is rejected by the admission webhook with vNPU nor supported for multiple devices, and the Pod might not be created.
• HAMi-managed hard-slice workloads do not currently provide workload-level metrics. Whole-card monitoring uses the Ascend NPU Exporter path, while HAMi soft slicing uses the built-in Pod and container metrics provided by this component.
• Soft slicing provides Pod and container memory metrics. Its compute utilization metric reports the utilization of the underlying physical device, so workloads sharing that device report the same value.
• Hard slicing requires driver support for virtual-device creation, and soft slicing requires driver support for device sharing. Use whole-card mode when the required driver capability is unavailable.
• With the v1.4.0 delivery default spec.config.create: false, changing only spec.hamiVnpuCore.enabled does not update the reused HAMi-owned hami-scheduler-device ConfigMap. Configure the intended target node in spec.nodeConfig; see Configure Ascend Slicing Mode.
• Ascend Driver 25.5 or later is required. The CANN libraries in the workload image must also match the installed Driver and, for soft slicing, the injected runtime ABI.
• The CANN ABI compatibility constraints are specific to the soft-slice path and do not define a complete model-framework support matrix; see Compatibility.

#v1.3.0

Alauda Build of HAMi Ascend Device Plugin v1.3.0 is a new operator-delivered device-plugin component for HAMi-managed Ascend allocation scenarios.

#New features and improvements

• Added Alauda Build of HAMi Ascend Device Plugin for HAMi-managed Ascend workloads.
• Added support for whole-card Ascend allocation through the HAMi-managed allocation path.
• Added vNPU hard-slice scheduling with HAMi scheduler device templates.
• Added Ascend vNPU soft-slice behavior through the HAMi Ascend vNPU runtime path.
• Exposes model-specific Ascend count and memory resources for HAMi-managed workloads.
• Use only resource names listed by the installed HAMi and HAMi Ascend Device Plugin package combination. Do not infer resource names from the chip model alone.

#Operational notes

• Stop the native Ascend Device Plugin (ascend-device-plugin) before enabling Alauda Build of HAMi Ascend Device Plugin on the same nodes.
• After changing device-plugin ownership on a node, confirm that the expected Ascend vNPU resources are allocatable before running workloads.
• For soft slicing, enable hamiVnpuCore for the target node and add huawei.com/vnpu-mode: hami-core to the workload. A workload without the annotation follows the template-based vNPU path.
• With the Ascend 310P resource configuration delivered in v1.3.0, a memory-only request may be normalized to the vir01 3072 MiB hard-slice template instead of receiving the exact requested quota.