Versions and Components

This page lists the HAMi components that users need to consider when installing, upgrading, or troubleshooting Alauda AI Infrastructure Accelerator Virtualization.

Current versions

ComponentTypeMarketplace versionPurpose
Alauda AI Infrastructure Accelerator VirtualizationOLM Operatorv2.10.0HAMi sharing, virtualization, and scheduling product.
Alauda AI Infrastructure Accelerator Virtualization ConsoleOperator Bundlev1.3.0Provides an optional HAMi-specific resource overview through the Helm Operator.
Alauda Build of HAMi Ascend Device PluginOperator Bundlev1.4.1Deploys and manages HAMi Ascend whole-card allocation, hard slicing, and soft slicing.
WARNING

Alauda AI Infrastructure Accelerator Virtualization Console v1.3.0 is delivered as an Operator Bundle based on the HAMi-WebUI v1.3.0 application line. Confirm the HAMi core version and Prometheus integration before upgrading.

Starting with v1.3.0, the delivery form is an Operator Bundle and the product version follows the upstream community HAMi-WebUI version line. The Alauda package form, images, and installation resources remain governed by the Operator Bundle release package.

v1.3.0 changes the delivery form from the earlier Cluster Plugin to an Operator Bundle. It does not support an in-place upgrade from the earlier Cluster Plugin. Uninstall the older WebUI Cluster Plugin and confirm its resources are removed before installing v1.3.0 and creating the HamiWebUI custom resource.

The v1.10.x Cluster Plugin and the v1.3.0 Operator Bundle belong to separate version lineages. The apparent numeric rollback reflects the delivery-form change and does not indicate a downgrade of the WebUI application.

The marketplace version and the upstream application baseline are separate version identifiers; select the marketplace version when installing or upgrading the package.

Alauda Build of HAMi Ascend Device Plugin v1.4.1 is the current release package. Review the release notes before use.

Ascend paths build on Alauda Build of NPU Operator for the Driver, OCI Runtime, and ascend RuntimeClass. Direct allocation uses its native Device Plugin; HAMi-managed allocation switches device exposure to HAMi Ascend Device Plugin. Use Compatibility for the supported product combination and environment boundaries, then Choose a HAMi Deployment Scenario before installing components.

Install the core HAMi Operator first. Add only the separately packaged components required by the selected backend and user tasks:

Use this table as the version source of truth. Use each installation page for its package form, prerequisites, and verification procedure.

Product components

ComponentPurpose
HAMi schedulerSchedules workloads that request HAMi resources.
HAMi NVIDIA device pluginManaged by the HAMi Operator through the Hami custom resource; reports NVIDIA resources to Kubernetes and maps selected NVIDIA devices into containers when enabled.
HAMi Ascend Device PluginProvides Ascend device exposure and allocation for HAMi-managed workloads, including whole-card allocation, template-based hard slicing, and soft slicing.
HAMi webhookAdjusts workload scheduling behavior for HAMi-managed resources.
HAMi-CoreProvides in-container resource control for supported backends.
HAMi WebUIProvides an optional HAMi-specific resource overview. It is installed as a separate cluster plugin when delivered by the target release.

Metrics prerequisites

Vendor metrics exporters are not HAMi components and are not versioned in this table. HAMi Ascend Device Plugin v1.4.1 also provides built-in metrics for soft slicing. See Monitor HAMi Usage for the metrics available through each backend path.

Image delivery boundaries

The v2.10.0 Operator package owns the hami-operator-bundle, hami-operator, and hami images. Choose the AMD64, ARM64, or multi-architecture package that matches the target environment. After installation, verify the running Operator, scheduler, and device-plugin Pod image IDs against the release package and the target cluster's registry mapping.

The Kubernetes scheduler image, NVIDIA driver and runtime, and workload CUDA image are supplied by the target cluster or workload. They are not HAMi package-owned related images. Confirm their architecture and acquisition path separately; a successful HAMi package import does not prove these external images are available.

ACP integration points

IntegrationValue
Operator namespacehami-system for the Subscription, CSV, Operator, and Hami custom resource
Runtime namespacekube-system for the HAMi scheduler and device-plugin components
Common NVIDIA resource keysnvidia.com/gpualloc, nvidia.com/gpucores, nvidia.com/gpumem
Common node labelsNVIDIA backend: gpu=on; Ascend backends: ascend=on
Ascend workload RuntimeClassascend for HAMi-managed whole-card, hard-slice, and soft-slice workloads
MonitoringACP dashboards, optional HAMi WebUI, and backend metrics components documented by the selected vendor base path
Quota metadataConfigured through ACP accelerator resource metadata
Optional WebUIAlauda AI Infrastructure Accelerator Virtualization Console Operator Bundle

This page lists combinations supported by Alauda delivery. Use upstream HAMi and vendor documentation for the broader backend and hardware matrix.

Lifecycle tasks