Conclusion

Module 1 — Deploy a Model: Deployed the open-source model for inference via KServe/vLLM on the GPU node, configured the model server runtime, replicas, and resource limits, and verified the deployment was accepting inference requests.

Module 2 — Evaluate with EvalHub: Ran a benchmark evaluation against the deployed model, analyzed latency, throughput, and quality results, and saved the evaluation as a repeatable workflow for tracking performance over time.

Module 3 — Model-as-a-Service: Exposed the model as a shared MaaS endpoint with subscription-based access control, an API key with token rate limits, and scaling policies for concurrent consumers.

Module 4 — Integrate into an Application: Configured and deployed the sample application against the MaaS endpoint using standard OpenAI-compatible API patterns.

Module 5 — Validate End-to-End: Traced a request through the serving chain (application → MaaS endpoint → model server → model), verified the response at each layer, and recovered from a simulated failure.

Module 6 — Monitor Performance: Generated load through the application and observed request rate, latency (p50/p95/p99), throughput, error rate, and resource utilization on the observability dashboard.

Red Hat AI Enterprise documentation — the full product documentation set

Red Hat OpenShift AI 3.5 documentation — serving, governance, and observability guidance

Validated models by Red Hat AI — third-party models validated for serving on the Red Hat AI platform

RedHatAI on Hugging Face — model details, weights, and deploy commands for validated models

Delete the API key you created in Module 3 so the subscription no longer grants access to the endpoint.

Delete the MaaS subscription and the model deployment from your project.

Delete the application deployed in Module 4.