The vLLM Connector
Inferra accelerates inference by serving KV-cache blocks that a serving engine would otherwise recompute. The component that connects the two is the Inferra vLLM connector (LcfVllmConnectorV1), a vLLM KV-connector plugin that runs inside the vLLM pod.
At a high level
vLLM asks the connector whether a prompt's prefix blocks are already cached.
On a hit, the connector loads them from the Inferra fabric instead of letting vLLM prefill them — which is where the time-to-first-token saving comes from.
On a miss, vLLM prefills normally and the connector stores the resulting blocks for the next request.
The connector is delivered as two Python wheels installed into the vLLM pod, not as part of this chart. It needs an Inferra cluster that is already deployed and healthy — everything up to and including Troubleshooting.
The full procedure is in the companion vLLM Connector Guide: which wheels to fetch from Cloudsmith, the environment variables and launch flags the vLLM engine needs, the pod-spec requirements, and how to confirm the connector loaded and is serving from cache. This page exists so you know what the connector is and where it fits; it does not repeat the procedure.
Two deployment properties that are easy to get wrong
These are properties of the deployment, not of the connector configuration, and getting either wrong produces a cluster that looks healthy and caches nothing.
A third is a property of the cluster rather than the pod, and it fails differently — quietly succeeding in the wrong place rather than not working: