1.7. Creating a ServingRuntime CR for use in MicroShift
Create a ServingRuntime custom resource (CR) based on installed manifests and release information. The included steps are an example of reusing the included microshift-ai-model-serving manifest files to re-create the OpenVINO Model Server (OVMS) model-serving runtime in the workload namespace.
This approach does not require a live node, so it can be part of CI/CD automation.
Prerequisites
-
Both the
microshift-ai-model-servingandmicroshift-ai-model-serving-release-infoRPMs are installed. - You have root user access to your machine.
-
The OpenShift CLI (
oc) is installed.
Procedure
Extract the image reference of the
ServingRuntimeCR you want to use from the MicroShift release information file by running the following command:$ OVMS_IMAGE="$(jq -r '.images | with_entries(select(.key == "ovms-image")) | .[]' /usr/share/microshift/release/release-ai-model-serving-"$(uname -i)".json)"1 - 1
- In this example, the image reference for the OVMS model-serving runtime is extracted.
Copy the original
ServingRuntimeYAML file by running the following command:$ cp /usr/lib/microshift/manifests.d/050-microshift-ai-model-serving-runtimes/ovms-kserve.yaml ./ovms-kserve.yamlAdd the actual image reference to the
image:parameter field value of theServingRuntimeYAML by running the following command:$ sed -i "s,image: ovms-image,image: ${OVMS_IMAGE}," ./ovms-kserve.yamlCreate the
ServingRuntimeobject in a custom namespace using the YAML file by running the following command:$ oc create -n <ai_demo> -f ./ovms-kserve.yaml1 - 1
- Replace
<ai_demo>with the name of your namespace.
If the ServingRuntime CR is part of a new manifest, set the namespace in the kustomization.yaml file, for example:
Example Kustomize manifest namespace value
apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
namespace: ai-demo
resources:
- ovms-kserve.yaml
#...
Next steps
-
Create the
InferenceServiceobject. - Verify that your model is ready for inferencing.
- Query the model.
- Optional: examine the model metrics.