AMD Prerequisites

Complete these steps before enabling AMD support.

Driver and cluster requirements

The cluster must provide:

• Kubernetes 1.34 or newer with the stable Dynamic Resource Allocation (DRA) API.
• Compatible AMD GPUs and a supported host kernel driver.
• The AMD Kubernetes operator with its DRA driver enabled.
• A gpu.amd.com DeviceClass and AMD devices published in ResourceSlices.
• Each published AMD device must include a non-empty string attribute named resource.kubernetes.io/pciBusID.

Validation used AMD GPU Operator v1.5.1 with a development build of the AMD DRA driver. This does not establish a minimum supported version. The installed DRA driver must meet the requirements listed above.

CM Platform does not install the AMD host driver or DRA driver.

Before enabling AMD support, check the DeviceClass and published devices:

kubectl get deviceclasses.resource.k8s.io gpu.amd.com
kubectl get resourceslices.resource.k8s.io \
  -o 'custom-columns=NAME:.metadata.name,DRIVER:.spec.driver,NODE:.spec.nodeName,DEVICES:.spec.devices[*].name'

Confirm that the DeviceClass exists and that gpu.amd.com rows list devices for each AMD node you intend to use.

The bundled cm-amd-vllm runtime uses seccompProfile: Unconfined. Your cluster's security policy must permit this container setting.

On OpenShift, the bundled security context constraints (SCCs) do not permit this setting. Before enabling AMD support, a cluster administrator must provide an SCC that permits it and grant the model engine's service account permission to use that SCC.

Prepare the nodes

The model agent uses this label to select AMD nodes for model caching:

kubectl label node <node-name> feature.node.kubernetes.io/amd-gpu=true

Apply the label to each AMD node that will serve models. Model deployments use the node and GPUs selected in the Portal.

Leave AMD nodes untainted or use the supported GPU taint:

kubectl taint node <node-name> amd.com/gpu=true:NoSchedule

The bundled model agent and AMD runtime tolerate this taint. Additional custom NoSchedule taints are not supported by the bundled configuration.

Prepare the cache on each selected node as described in Model Cache Path.

Enable AMD support

Add this setting to your installation values:

accelerators:
  amd:
    enabled: true

Apply the values using the CM Platform Deployment Guide. The default is false.

Enabling AMD support installs the cm-amd-vllm runtime and enables model-agent placement on labeled AMD nodes.

Verify model-agent placement

After the Argo CD applications are synced and healthy, check the pods:

kubectl -n ome get pods -o wide

Each AMD node used for model deployment must run a model-agent pod. This pod downloads and caches the model files. If a pod is missing, check the node label and taints before deploying a model.

Continue with Deploying on an AMD GPU.

Disable AMD support

Stop or delete all AMD deployments before setting accelerators.amd.enabled to false. Enable AMD support again before starting an AMD deployment.