|Index|Creating SUSE AI Factory Blueprints|What are the prerequisites?
Applies to SUSE AI Factory

2 What are the prerequisites?

Before creating and registering custom SUSE AI Factory blueprints, ensure your workstation and target cluster satisfy the necessary prerequisites.

2.1 Workstation requirements

To author, register and inspect blueprints, you need the Kubernetes command-line tool installed and configured with administrative access to your cluster:

  • Kubernetes CLI (kubectl): Must be installed on your local workstation.

    > kubectl version --client
  • Git: Recommended for versioning blueprint manifests, DAG files, and documentation.

2.2 Cluster requirements

Cluster prerequisites depend on whether they are required for all blueprints or only for specific workload features.

2.2.1 Mandatory requirements

Every SUSE AI Factory cluster hosting custom blueprints requires:

  • SUSE AI Factory Operator: The operator manages the Blueprint and AIWorkload custom resource definitions (CRDs). Without it, the cluster cannot interpret blueprint manifests.

    Verify the CRD is registered:

    > kubectl get crd blueprints.ai-factory.suse.com
  • SUSE Application Collection repository (ClusterRepo): Supplies standard Helm charts packaged for SUSE AI Factory (such as Ollama, vLLM, Apache Airflow, PostgreSQL, and Grafana).

    Verify the repository is registered:

    > kubectl get clusterrepo application-collection
  • Repository pull credentials: An Opaque Kubernetes secret containing authentication keys (user and token) so Helm can pull charts and container images from SUSE Application Collection.

    Verify the secret exists in the target operator namespace (typically aif-operator):

    > kubectl -n aif-operator \
      get secret application-collection \
      -o jsonpath='{.data}' | head -c 200

2.2.2 Conditional requirements

Depending on the charts and features declared in your blueprint, the cluster may require additional resources:

  • Default StorageClass (required for persistent components): Blueprints deploying databases (such as PostgreSQL or Qdrant) or caching models on disk require persistent volume claims (PVCs). If chart values do not specify an explicit storage class name, a default StorageClass must be defined.

    Verify that a default storage class is configured:

    > kubectl get storageclass | grep '(default)'
  • cert-manager (required for TLS Ingress): Needed if any blueprint component configures an external Ingress with automatic TLS certificates (for example, the Apache Airflow Web user interface).

    Verify that the cert-manager CRDs exist:

    > kubectl get crd certificates.cert-manager.io
  • NVIDIA GPU support (required for GPU inference engines like vLLM): Workloads serving models via vLLM require physical NVIDIA GPUs and the NVIDIA container runtime.

    Verify allocatable GPU capacity and the nvidia RuntimeClass:

    > kubectl get nodes \
      -o custom-columns=NAME:.metadata.name,GPU:.status.allocatable.nvidia\\.com/gpu
    > kubectl get runtimeclass nvidia
    Note
    Note

    Simulated GPU tools (such as kwok or fake-gpu-operator) can schedule GPU workload pods, but functional model inference requires physical GPUs.