Install OpenShift AI¶
Red Hat OpenShift AI is the MLOps layer on top of OpenShift. The tutorial uses it for KServe-based model serving (vLLM as the runtime). Other OpenShift AI components — workbenches, pipelines, distributed training — aren't required for this tutorial but won't hurt if they're enabled.
Path B users
If you're running on Developer Sandbox or CRC and using an external LLM endpoint, skip this guide. You don't need OpenShift AI installed.
Prerequisites¶
- An OpenShift 4.20+ cluster (see Choosing a Cluster)
cluster-adminrightsoclogged in to the cluster
This tutorial targets Red Hat OpenShift AI 3.x via the fast-3.x
channel (validated on 3.3.1). RHOAI 3.x requires OpenShift 4.19 or later.
Install the operator¶
From the OpenShift web console:
- Navigate to Operators → OperatorHub.
- Search for Red Hat OpenShift AI.
- Click the tile, then Install.
- Choose the
fast-3.xchannel (3.2 or later). - Accept the defaults (installed into
redhat-ods-operator).
Or from the CLI:
oc apply -f - <<EOF
apiVersion: operators.coreos.com/v1
kind: OperatorGroup
metadata:
name: rhods-operator
namespace: redhat-ods-operator
---
apiVersion: operators.coreos.com/v1alpha1
kind: Subscription
metadata:
name: rhods-operator
namespace: redhat-ods-operator
spec:
channel: fast-3.x
name: rhods-operator
source: redhat-operators
sourceNamespace: openshift-marketplace
EOF
Wait until the operator pods are running:
Create a DataScienceCluster¶
The DataScienceCluster (DSC) custom resource tells the operator which
components to enable. For this tutorial you need kserve managed; leave
the rest at the operator's defaults. Two non-obvious choices below get
inline comments.
apiVersion: datasciencecluster.opendatahub.io/v1
kind: DataScienceCluster
metadata:
name: default-dsc
spec:
components:
kserve:
managementState: Managed
# The rest of the tutorial assumes KServe Raw with a Headless predictor
# service — that's what produces the `:8000` URL caveat in serve-an-llm.md
# and install-ogx.md. RHOAI 3.x defaults to Headless; setting it
# explicitly documents the dependency.
rawDeploymentServiceConfig: Headless
dashboard:
managementState: Managed
llamastackoperator:
# RHOAI 3.x bundles a LlamaStack/OGX operator. We install the upstream
# ogx-k8s-operator in install-ogx.md instead (the rebrand hasn't shipped
# via RHOAI yet) — leaving this Removed avoids two operators reconciling
# the same LlamaStackDistribution.
managementState: Removed
Apply it:
Verify¶
The Dashboard hostname should also resolve. RHOAI 3.x exposes the
dashboard via Gateway API (not a plain Route in redhat-ods-applications):
If kserve shows Ready in the DSC status, you're done.