Other[AIW] Your first disaggregated deployment with NVIDIA Dynamo
This workshop will guide you through deploying your first disaggregated inference service with NVIDIA Dynamo on Oracle Kubernetes Engine (OKE). You will serve the Qwen3-0.6B model using vLLM, separate prompt processing from token generation, and send streaming inference requests to your deployment. Oracle provides the prepared OKE environment, workstation, cluster access, and GPU allocation. You will install the Dynamo operator with Helm and deploy three serving components: : Accepts API requests, routes work to the inference workers, and streams responses back to the client. : Processes the input prompt on one GPU and produces the attention state known as the key/value (KV) cache. : Uses a second GPU and the KV cache transferred through NIXL to generate output tokens. Both workers run on the same assigned node and each loads the model. The exercise uses two NVIDIA GPUs to demonstrate how prefill and decode work together. This workshop is ideal for developers and data scientists interested in: : Learn how separating prompt processing from token generation changes the structure of an inference deployment. : Configure worker placement and GPU requests, and understand the roles of namespaces, pods, and the Dynamo operator. : Deploy a model, inspect its health, and interact with its streaming API using command-line tools. By the end of this workshop, you will have hands-on experience with: : Explain the roles of the frontend, prefill worker, decode worker, vLLM inference engine, and NIXL KV-cache transfer. : Install the Dynamo operator with Helm, inspect a DynamoGraphDeployment manifest, and deploy Qwen3-0.6B with one GPU per worker using kubectl. : Verify pod readiness and placement, connect to the frontend with port-forwarding, and use curl to send prompts and inspect streamed responses. : Inspect pods, logs, and events, recognize common deployment issues, and remove your lab resources to release the allocated GPUs. Alejandro Casas, Sr. Principal Product Marketing Manager - OCI. Felipe Garcia, Solution Architect - NVIDIA. Having an issue with the lab? Have an idea on how it could be made better? Want to share your experience? Click the icon below to contact the team and let us know your feedback. Alejandro Casas - Senior Principal Product Marketing - October 2026.