Online installation from Helm
Install a new HAMi release with the official Helm chart. For an existing release, follow Upgrade.
Before installation
Complete the cluster and device prerequisites. Prepare the driver, container runtime, node labels, and values required by the device guide. For NVIDIA, save the applicable runtime and GPU Operator settings as hami-nvidia-values.yaml. Other vendors may require a separately deployed device plugin; enabling a vendor in HAMi does not install that vendor's plugin unless the chart provides it.
Add the Helm repository
helm repo add hami-charts https://project-hami.github.io/HAMi/
helm repo update hami-charts
helm search repo hami-charts/hami --versions
Select a chart version and use the matching version of this documentation. Next describes the development branch and may include settings absent from published charts.
Prepare values
Inspect the defaults for the selected version:
helm show values hami-charts/hami --version <chart-version>
Create my-values.yaml containing the overrides needed for the installation. For NVIDIA, start with the file prepared in the prerequisites guide. See Configuration for sharing limits, scheduling policies, and configuration precedence.
Image sources for networks in China
The chart uses upstream image registries by default (global.imageSource=upstream). If pulls from these registries are slow or fail on networks in China, set global.imageSource=cn. Check that the selected chart version includes global.imageSource in the helm show values output above before using it.
The DaoCloud public mirror (m.daocloud.io) may experience congestion or rate limiting. Selecting cn does not guarantee faster or successful image pulls. If pulls still fail, use an accessible upstream registry or copy the required images to a registry you control and configure the corresponding image registry and repository values.
To use global.imageSource=cn, keep global.imageRegistry empty (the default). A non-empty global.imageRegistry takes precedence and disables image-source selection.
Merge the following into my-values.yaml, preserving any other global settings:
global:
imageSource: cn
Alternatively, append --set global.imageSource=cn to the installation command below.
For supported default image references, cn selects Alibaba Cloud mirrors for kube-scheduler and the default pinned certgen image, and DaoCloud mirrors for HAMi, mock-device-plugin, and Lupine server/client images. Custom repositories are not automatically rewritten. See the Chart values reference for image overrides and exceptions.
Deploy HAMi
These examples use release hami in namespace kube-system. Substitute the actual release and namespace where needed. Replace <chart-version> with the version selected above:
helm install hami hami-charts/hami \
--namespace kube-system \
--version <chart-version> \
--values my-values.yaml
The chart detects the Kubernetes server version and selects the matching kube-scheduler image. To override the tag, set scheduler.kubeScheduler.image.tag in the values file. This image does not use global.imageTag.
NVIDIA driver discovery
Keep devicePlugin.nvidiaDriverRoot=auto for automatic discovery. The plugin reads GPU Operator's /run/nvidia/validations/driver-ready; if the file is absent, it assumes the driver is installed on the host. If HAMi starts before GPU Operator validation completes, wait for GPU Operator to become ready and restart the HAMi device plugin DaemonSet.
The device plugin Pod must use the NVIDIA runtime to load driver libraries and discover GPUs. Set NVIDIA as the node's default runtime or configure devices.nvidia.runtimeClassName. For CDI settings and prerequisites, see Configure CDI. Driver-path overrides and their mount behavior are documented in the Chart values reference.
Verify installation
The commands below use release hami in namespace kube-system with the default resource names. For a custom installation, use the actual namespace, release label, and resource names rendered by the chart. For example, release gpu with the default naming settings creates gpu-hami-scheduler and gpu-hami-device-plugin.
Check the scheduler and the device plugin Pods for the selected devices:
kubectl get pods -n kube-system -l app.kubernetes.io/instance=hami
kubectl rollout status deployment/hami-scheduler -n kube-system
For the chart-managed NVIDIA device plugin, also check:
kubectl rollout status daemonset/hami-device-plugin -n kube-system
A ready deployment confirms component startup. Run the NVIDIA example or the verification steps in the device guide to check allocation and workload behavior. Device plugins installed separately have their own resource names and namespaces.
For subsequent changes, use Upgrade. See Configuration for configuration options, or Troubleshooting for failures.
