Installing OpenShift AI Operator

Follow this procedure to install OpenShift AI 3.x.

  1. Create the namespace, operator group, and operator subscription. The stable-3.x channel tracks the latest 3.x release, 3.5.1 at the time of validation.

    apiVersion: v1
    kind: Namespace
    metadata:
      name: redhat-ods-operator
    ---
    apiVersion: operators.coreos.com/v1
    kind: OperatorGroup
    metadata:
      name: rhods-operator
      namespace: redhat-ods-operator
    spec: {}
    ---
    apiVersion: operators.coreos.com/v1alpha1
    kind: Subscription
    metadata:
      name: rhods-operator
      namespace: redhat-ods-operator
    spec:
      channel: stable-3.x
      name: rhods-operator
      source: redhat-operators
      sourceNamespace: openshift-marketplace
      installPlanApproval: Automatic
    

    Check which channels your catalog offers before you commit to one:

    oc get packagemanifest rhods-operator -n openshift-marketplace \
      -o jsonpath='{range .status.channels[*]}{.name}{"="}{.currentCSVDesc.version}{"\n"}{end}'
    

    Channel

    Use for

    stable-3.x

    Latest 3.x release. The default channel and the recommended choice for new installations.

    stable-3.5

    Pins to one minor release so upgrades do not move you across minor versions.

    eus-3.5

    Extended Update Support for that minor release.

    beta

    Pre-release builds. Not for production.

    Important

    Always name a 3.x or 3.y channel explicitly. The unsuffixed stable, fast, and alpha channels still track an older release train, so subscribing to them installs a release that this chapter does not describe.

  2. Verify that the operator is installed.

    oc get csv -n redhat-ods-operator
    oc get pods -n redhat-ods-operator
    
  3. Confirm that DSCInitialization reconciled. The operator creates default-dsci automatically and it reaches Ready without any input.

    oc get dsci
    
    NAME           AGE   PHASE   CREATED AT
    default-dsci   71s   Ready   2026-09-24T11:30:18Z
    

    The resource uses dscinitialization.opendatahub.io/v2, whose spec has only applicationsNamespace, monitoring, trustedCABundle, and devFlags. There is no service mesh configuration to supply. Customize it only if you need to change the applications namespace or monitoring:

    apiVersion: dscinitialization.opendatahub.io/v2
    kind: DSCInitialization
    metadata:
      name: default-dsci
    spec:
      applicationsNamespace: redhat-ods-applications
      monitoring:
        managementState: Managed
        namespace: redhat-ods-monitoring
      trustedCABundle:
        managementState: Managed
    
  4. Create the DataScienceCluster resource to enable the required RHOAI components. This example enables the minimum needed for GPU-accelerated model serving and notebooks.

    apiVersion: datasciencecluster.opendatahub.io/v2
    kind: DataScienceCluster
    metadata:
      name: default-dsc
    spec:
      components:
        dashboard:
          managementState: Managed
        kserve:
          managementState: Managed
        workbenches:
          managementState: Managed
          workbenchNamespace: rhods-notebooks
        aipipelines:
          managementState: Removed
        feastoperator:
          managementState: Removed
        kueue:
          managementState: Removed
        llamastackoperator:
          managementState: Removed
        mlflowoperator:
          managementState: Removed
        modelregistry:
          managementState: Removed
        ogx:
          managementState: Removed
        ray:
          managementState: Removed
        sparkoperator:
          managementState: Removed
        trainer:
          managementState: Removed
        trainingoperator:
          managementState: Removed
        trustyai:
          managementState: Removed
    

    List the full set of components your release supports with:

    oc get crd datascienceclusters.datasciencecluster.opendatahub.io -o json | \
      jq -r '.spec.versions[]|select(.name=="v2")|.schema.openAPIV3Schema.properties.spec.properties.components.properties|keys[]'
    

    Note

    The kserve component takes no serving stanza, because KServe runs in RawDeployment mode. Its fields cover RawDeployment behavior, the OAuth proxy sidecar, model caching, and the LLMInferenceService feature flags.

    To enable queue-based scheduling, install the Red Hat build of Kueue Operator and set kueue.managementState: Unmanaged so that the Kueue Operator owns the Kueue installation. Then set ray and trainingoperator to Managed as needed.

  5. Apply the manifest and wait for the cluster to become ready.

    oc apply -f dsc.yaml
    oc wait --for=condition=Ready datasciencecluster/default-dsc --timeout=900s
    
  6. Verify the deployed components.

    oc get dsc
    oc get pods -n redhat-ods-applications
    

    Review the per-component conditions if the cluster does not reach Ready:

    oc get dsc default-dsc -o json | jq -r '.status.conditions[] | "\(.type)=\(.status) \(.reason): \(.message)"'
    

    Conditions reporting reason: Removed are expected for the components you disabled. Conditions such as KserveLLMInferenceServiceDependencies=False PreConditionFailed: Red Hat Connectivity Link not installed are informational and do not block Ready; they only matter if you intend to use distributed inference with llm-d.

  7. Find the dashboard URL. The dashboard is published through a Gateway API Gateway.

    oc get gateway -n openshift-ingress
    oc get route data-science-gateway -n openshift-ingress -o jsonpath='{.spec.host}{"\n"}'
    

    The dashboard is served at https://rh-ai.apps.<cluster-domain>, which is the host on the data-science-gateway route in openshift-ingress. Two further routes, rhods-dashboard and a second data-science-gateway, exist in redhat-ods-applications and redirect to that host, so older bookmarks keep working.