NVIDIA published a walkthrough on 30 October 2026 describing how NVIDIA AI Cluster Runtime (AICR) packages Dynamo and its Kubernetes infrastructure into versioned recipes. The recipes record pinned component versions, compatibility constraints, configuration values, checksums, and validation expectations, giving operators a defined configuration against which to check a deployment.[1]

The article specifies the Dynamo Platform Helm chart at version 1.4.2. Its example AICR v1.0.0 recipe targets Amazon EKS on NVIDIA GB200 NVL72 and GB300 NVL72 systems. NVIDIA says its team tested the listed component versions together on those platforms and hardware; that claim should not be generalized to other environments.[1]

The workflow snapshots a cluster, generates a recipe, renders a deployment bundle, installs the runtime, and validates the live cluster before application deployment. Depending on the recipe, deployment and conformance checks cover GPU capacity, operator health and versions, accelerator access, scheduling, metrics, and ingress readiness. Offline validation evaluates declarative constraints and skips behavioral checks. Performance validation runs separately and consumes GPUs through a temporary serving workload.[1]

The walkthrough also distinguishes Dynamo Frontend routing from Gateway API Inference Extension routing with an Endpoint Picker Plugin. Both use worker-reported key-value (KV) cache events in the illustrated configuration. NVIDIA warns that predicting cache contents from prior routing decisions cannot observe evictions and can let the router’s view drift.[1]

[1]: NVIDIA Technical Blog, “Build a Validated Kubernetes Runtime for NVIDIA Dynamo,” 30 October 2026. https://developer.nvidia.com/blog/build-a-validated-kubernetes-runtime-for-nvidia-dynamo

Sources

  1. Build a Validated Kubernetes Runtime for NVIDIA Dynamodeveloper.nvidia.com · 30 Oct 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

7a90ebd638ef4bbe8951add6c4a2f2fd34536b05de3b6b6dfaf65899f4d4ba5b