Skip to main content
Version: 3.6.1-rc.2

Intel Prerequisites

Complete these steps before enabling Intel support.

Driver and cluster requirements​

The cluster must provide:

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

Validation used Intel Arc Pro B70 GPUs with the Intel GPU base operator 0.4.0. That operator installed the Intel GPU resource driver v0.12.0 and the xpumd v2.2.0 service. This does not establish a minimum supported version. The installed DRA driver must meet the requirements listed above.

The xpumd service publishes each GPU's memory size. Without it, the Portal shows the GPUs without their memory size.

The Portal reads the memory use and the activity of each Intel GPU from the xe kernel driver. On a node that uses the older i915 driver, the Portal lists the GPUs but shows usage unavailable for them and draws no line for them in Recent GPU activity.

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

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

kubectl get deviceclasses.resource.k8s.io gpu.intel.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.intel.com rows list devices for each Intel node you intend to use.

Do not also give these GPUs to workloads through the Intel GPU device plugin (gpu.intel.com/xe or gpu.intel.com/i915 resources). The bundled Intel runtime gets its GPUs only through DRA.

Prepare the nodes​

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

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

When the Intel operator uses Node Feature Discovery (NFD) labeling, NFD sets this label for you. Check that each Intel node that will serve models has it. Model deployments use the node and GPUs selected in the CM portal.

Leave Intel nodes untainted or use the supported GPU taint:

kubectl taint node <node-name> intel.com/gpu=present:NoSchedule

The bundled model agent, Intel runtime, and platform log collector tolerate this taint with any value. When workload logging is enabled, the log collector can collect model logs from tainted Intel nodes for the Portal's Logs view.

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 Intel support​

Add this setting to your installation values:

accelerators:
intel:
enabled: true

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

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

It also runs the GPU memory collector cm-purplepill-drm in the monitoring namespace, one pod on each labeled Intel node. The collector reads GPU memory from the kernel and needs no Intel tools. It uses the host PID namespace and a privileged container, so the security policy of the monitoring namespace must allow both. On OpenShift, the bundled cm-purplepill security context constraint allows them. To check the collector, run this command. Each labeled node runs one pod.

kubectl -n monitoring get pods -l app=cm-purplepill-drm -o wide

Verify model-agent placement​

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

kubectl -n ome get pods -o wide

Each Intel 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 or Intel GPU.

Disable Intel support​

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