Inferra vLLM Connector Guide

This guide is intended for operators who already have an Inferra LCF cluster running and want to attach the LCF KV-cache connector to a vLLM serving pod. It covers only how to install and wire the connector plugin into a vLLM engine.

Standing up the cluster is not covered here

Standing up the LCF cluster itself (LCFC/LCFT, lcf-api, node labeling, XFS mounts, Cloudsmith login) is not covered here — see the Deployment and Usage Guide, which documents the lcf-cli deploy and the direct-Helm deploy. This guide assumes that one has been followed and the cluster is healthy.

What is in this guide

Values to Substitute Before Running

Placeholder

Meaning

x.y.z

Connector / pylcf release version on Cloudsmith. Provided by Lightbits Support.

<vllm-namespace>

The Kubernetes namespace the vLLM pod runs in (default vllm). All kubectl commands in this guide use this.

<vllm-deployment>

The name of your vLLM Deployment (for example vllm-lcf-qwen).

<lcf-cache-scope>

The connector's logical KV-cache scope label (the lcf_namespace config value). This is not a Kubernetes namespace — it is a tag the connector stamps on its cached KV. You pick this value, and it must name a logical namespace that exists in the cluster. Pods that share cached KV use the same value: several vLLM replicas of the same model can bind to one scope, and each reuses the KV the others stored — see Connector Environment and Launch Flags.

Which version to use for x.y.z

The connector/pylcf wheel version (x.y.z) tracks the LCF release version. The version in the current Deployment and Usage Guide is 1.2.0 (PEP 440: 1.2.0) — so unless Lightbits Support tells you otherwise, use x.y.z = 1.2.0.