Skip to main content

HAMi

The HAMi addon provides GPU virtualization: it lets several pods share the same GPU, instead of assigning a whole GPU to each pod.

In the console​

HAMi requires the GPU Operator addon, and therefore a node group with GPUs.

  1. At creation, Addons step, check HAMi (disabled by default). If GPU Operator is not checked, the console shows "Requires the GPU Operator addon" and blocks the next step.
  2. On an existing cluster: Edit > Extensions & Addons, check HAMi, then Save.

The cluster detail page shows HAMi in the Extensions section when it is active.

Override the configuration​

Once the addon is checked, the Helm Configuration (YAML) — optional field appears. The value is passed to the HAMi Helm chart, under the hami key. The available options are described in the HAMi documentation.

Usage in the cluster​

# HAMi pods
kubectl get pods -A | grep -i hami

Pods request a fraction of a GPU using the resources exposed by HAMi (GPU memory, compute share). See the HAMi documentation for the resource names and pod examples.

Best practices​

  • Reserve GPU sharing for workloads that do not use a whole GPU (light inference, development, notebooks).
  • Set GPU memory limits per pod to prevent one pod from starving the others.