Install the Commercial HAMi NVIDIA vGPU Addon¶

The commercial HAMi NVIDIA vGPU Addon is used to virtualize an NVIDIA GPU into multiple vGPUs in a Kubernetes cluster and allocate them to different workloads on demand.

The installation procedure of the commercial edition is basically the same as that of the open-source NVIDIA vGPU Addon. The main difference is that after the commercial edition is installed, you need to import a valid License and complete activation before the related commercial capabilities can be used normally.

Differences Between the Commercial Edition and the Open-source Edition¶

Installation Differences¶

Item	Open-source Edition	Commercial Edition
Installation entry	Helm Chart installation	Helm Chart installation
Helm chart	`nvidia-vgpu`	`nvidia-vgpu-commercial`
Installation parameters	Basically the same	Basically the same
GPU mode switching	Supports switching to vGPU mode or MIG mode	Supports switching to vGPU mode or MIG mode
License	Not required	A License must be imported and activated

Feature Differences¶

Feature	Open-source Edition	Commercial Edition	Description
Heterogeneous card support	All NVIDIA	All NVIDIA; All PPU; Ascend 910A, 910B2, 910B3, 910B4, 910B4-1, 910C (slicing not supported), 310P; Metax Xiyun series GPU (including C550/C500/C500X/C290/C280/N260 and other models); Hygon; Cambricon; Enflame; Kunlunxin; Moore Threads; Iluvatar and more	—
vGPU resource slicing	Computing power slicing	Computing power slicing; Memory slicing	—
Scheduling capability	For NVIDIA GPU: Binpack/Spread Specify the GPU card model Specify a specific card	For NVIDIA GPU: Binpack/Spread Specify the GPU card model Specify a specific card Task priority	—
Monitoring capability	Supported	Supported	Can be integrated with the observability module via ServiceMonitor
Enterprise-level support	Open-source community support	Commercial support	—
License management	No License required	License activation required	—

Prerequisites¶

Before installation, confirm the following:

Refer to the GPU Support Matrix to confirm that the cluster nodes have NVIDIA GPU cards of the corresponding model.
The GPU Operator has been deployed in the current cluster through a Helm app. For details, refer to Offline Install GPU Operator.
For the commercial HAMi, you need to apply for a License from HAMi in advance.

Installation Steps¶

The installation method of the commercial Addon is the same as that of the open-source edition. The difference lies in the Helm chart name and the License activation after installation.

Access the target cluster.

Path: Container Management -> Clusters -> click the target cluster name.
Go to the Helm Chart page and select the commercial Addon.

Path: Helm Apps -> Helm Chart , search for and select the Addon corresponding to nvidia-vgpu-commercial.

Configure the installation parameters.

The common parameters are as follows:

Parameter	Description
`deviceCoreScaling`	The GPU computing power usage ratio, with a default value of 1. A value greater than 1 indicates that virtual computing power is enabled. If configured as S, the total computing power of the vGPUs sliced from a single GPU is S × 100%.
`deviceMemoryScaling`	The GPU memory usage ratio, with a default value of 1. A value greater than 1 indicates that virtual memory is enabled. If the physical memory of the GPU is M, when configured as S, the total memory of the sliced vGPUs is S × M.
`deviceSplitCount`	The maximum number of tasks that can be sliced from a single GPU, with a default value of 10. At most N tasks can exist simultaneously on each GPU (where N is the value of this parameter).
`Resources`	The resource requests and limits of components such as vgpu-device-plugin and vgpu-scheduler.
`ServiceMonitor`	Disabled by default. Once enabled, you can go to the observability module to view vGPU-related monitoring. If you need to enable it, make sure that insight-agent is installed and running; otherwise, the NVIDIA vGPU Addon installation will fail.

To modify advanced parameters, you can edit them directly in the YAML column.

Submit the installation.
Confirm that the Addon-related Pods are running properly.
```
kubectl get pod -n <the namespace where HAMi is located>
```
Expected result: Pods of components such as the HAMi scheduler and device plugin are all in the Running state.
Switch the node GPU mode to vGPU.

Click Nodes in the left navigation bar, find the target node, click Switch GPU Mode , and switch to vGPU mode.

Note

NVIDIA's vGPU capability supports node-level GPU mode switching (full GPU/vGPU/MIG mode), meeting the different GPU mode requirements of different workloads in the same cluster.

After clicking Confirm , the node status changes to Switching GPU Mode . After the switch is complete (that is, the hami-nvidia-vgpu-device-plugin Pod for vGPU has started up), the node status changes to Nvidia-vGPU .

After the node GPU mode is switched successfully, you can refer to Use NVIDIA vGPU in Applications to deploy workloads. The switching process has a slight delay, so deploy applications only after the node labels are displayed correctly.

Import a License and Activate It¶

After the commercial HAMi is installed, you must import a License to activate the commercial capabilities.

Obtain the License Application Information¶

Before or after installation, you need to prepare the License application information (such as the GPU UUID) as required by HAMi. For the detailed steps to obtain the application information, import the License, and complete activation, refer to Obtain the Commercial HAMi License Information and Complete the Import.

Verify the Installation¶

After completing the installation, GPU mode switching, and License activation, you can verify the installation in the following ways:

In Nodes , confirm that the GPU mode of the target node is Nvidia-vGPU .
Run kubectl get pod -n <the namespace where HAMi is located> and confirm that all related components are in the Running state.
Deploy a test workload and confirm that resources such as nvidia.com/vgpu can be requested normally.