#Install Alauda Build of HAMi

This page describes how to install the core Alauda Build of HAMi cluster plugin in ACP. HAMi-WebUI and the HAMi Ascend Device Plugin are separately packaged components with their own installation tasks.

#Before you begin

• Cluster administrator access to the target ACP cluster.
• An Alauda Build of HAMi package supported by the target ACP and Kubernetes versions. See Versions and Components.
• Accelerator nodes have the required vendor driver, runtime, and node readiness checks completed through the corresponding NVIDIA GPU vendor base documentation or Ascend NPU vendor base documentation.
• The target backend is supported by the current HAMi package and ACP release.

#Choose a deployment scenario

Before installing HAMi, use Choose a HAMi Deployment Scenario to determine the required products, HAMi settings, installation sequence, and device exposure owner. Return to this page after selecting the scenario.

Select one vendor backend for the target cluster. This procedure does not cover enabling NVIDIA GPU and Ascend NPU backends in the same cluster.

#Prepare the cluster plugin

Retrieve Alauda Build of HAMi from Customer Portal and upload the cluster plugin package to ACP. Please contact Consumer Support for package access. For the standard upload procedure, see Upload Packages.

#Install Alauda Build of HAMi

1. Label the nodes that should run HAMi components. Use the node selector required by the selected backend and the plugin form.

   For common NVIDIA backend examples:

   kubectl label nodes <node-name> gpu=on

   For common Ascend backend examples:

   kubectl label nodes <node-name> ascend=on
   WARNING

   Keep one device exposure and resource allocation owner for the same physical devices.

   If HAMi exposes devices on a node, do not also let a vendor device plugin or DRA driver expose the same devices on that node. For Ascend, use Switch the Ascend Device Plugin Owner for the maintenance procedure.

2. In ACP, go to Administrator -> Marketplace -> Cluster Plugins, switch to the target cluster, and deploy Alauda Build of HAMi.

3. Configure the plugin form for the selected backend and runtime.

   Field	How to decide
   Enable NVIDIA	Enable it when HAMi manages NVIDIA GPU nodes. It is enabled by default.
   NVIDIA Runtime Class Name	Use the RuntimeClass prepared for HAMi NVIDIA workloads. The plugin form commonly defaults to hami-nvidia.
   Create NVIDIA Runtime Class	Enable it only when the named RuntimeClass does not already have another owner.
   NVIDIA Device List Strategy	Use envvar for NVIDIA_VISIBLE_DEVICES injection. Use cdi-annotations only after the NVIDIA base path has prepared CDI and the driver and hook paths are known.
   NVIDIA Device Split Count	Maximum number of concurrent HAMi tasks allowed on one GPU. The default is 10.
   NVIDIA Device Memory Scaling	Memory overcommit ratio. Keep 1 unless overcommitment is part of the capacity plan.
   NVIDIA Device Core Scaling	Compute overcommit ratio. Keep 1 unless overcommitment is part of the capacity plan.
   Node Scheduler Policy	Choose binpack to consolidate workloads or spread to distribute them across nodes.
   GPU Scheduler Policy	Choose binpack to consolidate workloads or spread to distribute them across devices.
   Custom Kubernetes Scheduler Image	Leave empty to use the image generated from the cluster Kubernetes version and configured registry. Set a complete image reference only when the cluster requires an override.
   Enable Ascend	Enable it when HAMi schedules Ascend NPU or vNPU resources. It is disabled by default.

   NOTE

   • Enable NVIDIA enables the HAMi-built-in NVIDIA device plugin.
   • Enable Ascend enables Ascend support in the HAMi scheduler.
   • Alauda Build of HAMi Ascend Device Plugin separately exposes Ascend devices for HAMi-managed allocation.

   For overcommitment behavior, see Configure Overcommitment Ratio. For CDI preparation and driver paths, continue in the NVIDIA GPU vendor base documentation.

4. Verify that the HAMi scheduler is running:

   kubectl get pods -n kube-system | grep "hami-scheduler"

   For the NVIDIA scenario, also verify the built-in NVIDIA device plugin:

   kubectl get pods -n kube-system | grep "hami-device-plugin"

5. For the NVIDIA scenario, confirm that nodes expose the expected HAMi NVIDIA resource keys:

   kubectl get node <node-name> -o jsonpath='{.status.allocatable}'

   For an Ascend scenario, install Alauda Build of HAMi Ascend Device Plugin before checking node allocatable resources. The core HAMi installation does not expose Ascend devices by itself.

6. For the NVIDIA scenario, register ACP quota metadata for the HAMi resource names that users will request. For an Ascend scenario, complete the HAMi Ascend Device Plugin installation and confirm its resource names before registering the metadata. For the ACP-owned resource metadata procedure, see Accelerator Resource Quota.

#Continue with backend-specific components

• For HAMi-managed Ascend allocation or vNPU, continue with Install HAMi Ascend Device Plugin.
• For the optional HAMi resource WebUI, continue with Install HAMi-WebUI.

Direct Ascend NPU allocation without HAMi scheduling is outside this HAMi-managed installation path. Continue with Request Ascend NPU resources directly for that workload model.

#Next steps

• Verification
• Install HAMi Ascend Device Plugin
• Install HAMi-WebUI
• Choose a HAMi backend path
• Request HAMi resources
• Monitor HAMi usage