Configuring GPU Profiles

RHOAI needs a hardware profile to offer Intel GPUs to Data Science Projects, workbenches, and model deployments.

HardwareProfile uses the infrastructure.opendatahub.io/v1 API. The earlier v1alpha1 version is still served for compatibility but is no longer the storage version, so write new manifests against v1.

The v1 API adds a resourceType field on each identifier, which takes CPU, Memory, or Accelerator. Set it so the dashboard renders the right control for each resource.

Creating a hardware profile

  1. Inspect the profile RHOAI creates by default. It is a useful template for the CPU and memory identifiers every profile should carry.

    oc get hardwareprofile default-profile -n redhat-ods-applications -o yaml
    
  2. Choose the manifest that matches how your cluster exposes GPUs.

    • Use this profile if your cluster exposes the gpu.intel.com device class through DRA:

      apiVersion: infrastructure.opendatahub.io/v1
      kind: HardwareProfile
      metadata:
        name: intel-gpu-dra
        namespace: redhat-ods-applications
        annotations:
          opendatahub.io/display-name: "Intel GPU (DRA)"
          opendatahub.io/description: "Intel Arc Pro B-Series GPU exposed through Dynamic Resource Allocation"
      spec:
        identifiers:
        - displayName: CPU
          identifier: cpu
          resourceType: CPU
          minCount: 2
          maxCount: 8
          defaultCount: 2
        - displayName: Memory
          identifier: memory
          resourceType: Memory
          minCount: 4Gi
          maxCount: 32Gi
          defaultCount: 8Gi
        - displayName: Intel GPU (DRA)
          identifier: gpu.intel.com
          resourceType: Accelerator
          minCount: 1
          maxCount: 8
          defaultCount: 1
        scheduling:
          type: Node
          node:
            nodeSelector:
              intel.feature.node.kubernetes.io/gpu: "true"
      
    • Use this profile if your cluster exposes gpu.intel.com/xe through the Intel device plugin:

      apiVersion: infrastructure.opendatahub.io/v1
      kind: HardwareProfile
      metadata:
        name: intel-gpu-xe
        namespace: redhat-ods-applications
        annotations:
          opendatahub.io/display-name: "Intel GPU (device plugin)"
          opendatahub.io/description: "Intel Arc Pro B-Series GPU exposed through the Intel GPU device plugin"
      spec:
        identifiers:
        - displayName: CPU
          identifier: cpu
          resourceType: CPU
          minCount: 2
          maxCount: 8
          defaultCount: 2
        - displayName: Memory
          identifier: memory
          resourceType: Memory
          minCount: 4Gi
          maxCount: 32Gi
          defaultCount: 8Gi
        - displayName: Intel GPU
          identifier: gpu.intel.com/xe
          resourceType: Accelerator
          minCount: 1
          maxCount: 8
          defaultCount: 1
        scheduling:
          type: Node
          node:
            nodeSelector:
              intel.feature.node.kubernetes.io/gpu: "true"
      

    With DRA, the identifier is the DeviceClass name, gpu.intel.com, not a Kubernetes extended resource name.

    Each identifier requires displayName, identifier, minCount, and defaultCount. maxCount and resourceType are optional but recommended.

  3. Apply the selected profile and verify that RHOAI reports it.

    oc apply -f hardware-profile.yaml
    oc get hardwareprofile -n redhat-ods-applications
    
  4. Verify the profile appears in the RHOAI dashboard under Settings > Hardware profiles.

Scheduling options

spec.scheduling controls how workloads using the profile are placed.

Type

Behavior

Node

Applies the given nodeSelector and tolerations directly to the workload pod. Use this when you have no queueing layer.

Queue

Submits the workload to the Kueue LocalQueue named in spec.scheduling.kueue.localQueueName. The queue's configuration then determines node placement and tolerations. Requires the Red Hat build of Kueue Operator and kueue.managementState: Unmanaged in the DataScienceCluster.

Note

Do not set both a nodeSelector under scheduling.node and a Kueue local queue. The kueue stanza is only valid when the scheduling type is Queue.