OpenShift AI

What problem does it solve?

Workshop and Industrial Edge teams need model serving, notebooks/workbenches, and external LLM access without standing up inference infrastructure per user. On the hub, OpenShift AI (RHODS) provisions DataScienceCluster, per-user namespaces (ai-user1…), and MaaS proxies to RHDP LiteMaaS. On spokes, a lighter RawDeployment KServe stack scores anomaly models against Kafka telemetry.

NeuroFace (OVMS PPE Safety webcam) and Developer Hub Lightspeed use MaaS models (llama-scout-17b, granite-3-2-8b-instruct).

Hub component

Path Purpose
charts/all/openshift-ai-hub/ DataScienceCluster, per-user ai-userN + notebook neuroface-ml-lab, maas-workshop, MaaS proxies
charts/all/developer-hub/templates/catalog-openshift-ai.yaml Catalog System openshift-ai-workshop — one Component per user

Per-user projects (userN)

Pattern Description
Namespace ai-user1 … ai-user50 (opendatahub.io/dashboard: "true")
RBAC userN → admin on ai-userN
Workbench Notebook neuroface-ml-lab — microp learning + MaaS chat lab
MaaS Secret maas-connection per project
Shared maas-workshop — playground + MaaS InferenceService proxies

How to reach your project (userN)

  1. OpenShift Console → OpenShift AI (Platform Hub-Spoke menu)
    Or: https://rh-ai.apps.<hub-domain>
  2. Applications → Enabled → ensure dashboard is enabled
  3. Data Science Projects → select ai-userN (matches workshop login userN)
  4. Workbenches → start neuroface-ml-lab
  5. Cross-link: NeuroFace UI at https://neuroface.<hub-domain> for OVMS face-detection-retail-0005 (PPE Safety webcam)

Developer Hub: Catalog → System openshift-ai-workshop → Component ai-userN.

Model serving (hub workshop track)

Component DSC setting (chart default)
Workbenches workbenches.managementState: Managed
KServe RawDeployment — MaaS proxy InferenceServices in maas-workshop
ModelMesh / Serverless Optional — disabled by default (see below)

Do I need OpenShift Serverless for KServe?

No, for this pattern’s default. OpenShift AI 3.x + KServe can run in RawDeployment mode (modelServing.defaultDeploymentMode: RawDeployment, kserve.serving.managementState: Removed). InferenceServices deploy as normal Deployments — no Knative, no KnativeServing CR.

Install OpenShift Serverless only if you explicitly enable modelServing.serverlessEnabled: true (Knative-based autoscaling / scale-to-zero). The hub chart still lists serverless-operator in subscriptions for optional use; KServe on the hub does not depend on it with current values.

CPU models on MinIO (automated)

Step Chart What happens
1 industrial-edge-minio MinIO + bucket models (+ quay, kubecost)
2 industrial-edge-minio PostSync Job minio-model-seed uploads CPU sklearn model to s3://models/anomaly-detection/model/
3 openshift-ai-hub Secret aws-connection-models in maas-workshop and each ai-userN (OpenShift AI dashboard S3 connection)
4 openshift-ai-hub InferenceService workshop-sklearn (KServe RawDeployment) reads the MinIO model

Spokes reach the same bucket via Skupper minio-hub (http://minio-hub.service-interconnect.svc:9000) — see industrial-edge-data-science-project.

Console: https://minio-console-industrial-edge-ml-workspace.<hub-apps-domain> — browse/upload CPU model artifacts to bucket models.

MaaS models (external)

Endpoint: https://maas-rhdp.apps.maas.redhatworkshops.io/v1 — keys via RHDP litemaas.apiKey or Secret maas-facilitator-seed (Vault+ESO).

Model Use case
llama-scout-17b Default userN / NeuroFace chat
granite-3-2-8b-instruct Developer Hub Lightspeed
deepseek-r1-distill-qwen-14b Optional ODS connection
codellama-7b-instruct Code / templates

Console

  • ConsoleLink: Platform Hub-Spoke → OpenShift AI
  • URL: https://rh-ai.apps.<hub-apps-domain>
  • Legacy (2.x): https://rhods-dashboard-redhat-ods-applications.<hub-apps-domain>/ — OpenShift AI 3.4 operator redirects to rh-ai.apps.*; bookmarks may still use the old hostname during transition.

Developer Hub

  • Catalog: openshift-ai-workshop / maas-workshop-shared
  • Kuadrant LLM API: /kuadrant → workshop MaaS API Product

GPU inference (optional)

Default pattern path is CPU-only (RawDeployment KServe, OVMS openvino_ir on spokes). GPU accelerates OVMS face-detection, YOLO PPE, and enables self-hosted LLM (vLLM/TGIS) without external MaaS.

Prerequisites (install order)

  1. Node Feature Discovery (NFD) — redhat-operators, channel stable; create NodeFeatureDiscovery CR
  2. NVIDIA GPU Operator — certified-operators, channel stable; create ClusterPolicy CR (defaults OK for T4/A10G/A100)
  3. Wait for nvidia-driver-daemonset Ready on GPU nodes (~5 min)
  4. Verify: oc exec -n nvidia-gpu-operator <driver-pod> -- nvidia-smi
  5. OpenShift AI DataScienceCluster detects GPUs automatically in the dashboard

Optional: NVIDIA Network Operator (distributed training), OpenShift Serverless (Knative KServe), NVIDIA NIM Operator (enterprise LLM).

InferenceService with GPU

Add to predictor container resources when GPU nodes are labeled:

resources:
  limits:
    nvidia.com/gpu: "1"
Workload CPU path GPU path VRAM
OVMS face-detection ~1 CPU, ~2 GiB ~0.5 GPU ~2 GiB
YOLO PPE ~0.2–2 CPU ~0.5 GPU ~2 GiB
vLLM 7B (self-hosted) N/A (use MaaS) 1 GPU ~16 GiB
vLLM 14–70B N/A 1× A100 24–80 GiB

Sizing: see Bill of Materials — GPU. Verify: CHECK_GPU=1 ROLE=spoke bash scripts/verify-node-capacity.sh

Pattern charts do not install NFD/GPU Operator by default. GPU support in spoke-neuroface / neuroface charts is roadmap (optional gpu.enabled flag).

Documentation