What the Connector Is

LcfVllmConnectorV1 is a vLLM KV-cache connector plugin. It lets vLLM offload and re-load KV-cache blocks to/from the Lightbits LCF RDMA/TCP fabric instead of recomputing them.

The two wheels

Wheel

What it is

lcf-vllm-connector

The thin vLLM adapter. Registers LcfVllmConnectorV1.

pylcf

The compiled Rust engine it forwards to — hashing, fabric key-fold, gather, device copy, streaming prefetch, residence/completion.

Where it sits

┌───────────────────────────────────────────────────┐
│ vLLM pod (v0.30.0 default) │
│ Kubernetes namespace: vllm │
│ │
│ vLLM engine │
│ │ KV blocks (offload / reload) │
│ ▼ │
│ LcfVllmConnectorV1 ──► pylcf (Rust engine) │
└─────────────────────────┬─────────────────────────┘
│ Unix socket /var/run/lcf/lcfc.sock (node hostPath)
▼
┌─ Lightbits Inferra Cluster (LCF) ─────────────────┐
│ Kubernetes namespace: lcf │
│ │
│ ┌───────────────────────────────────────────┐ │
│ │ LCFC (node-local agent) │ │

The connector runs inside the vLLM pod (Kubernetes namespace vllm), while LCFC/LCFT run in the LCF cluster's own namespace (lcf). They are not connected over Kubernetes networking: the connector talks to the node-local LCFC over a host Unix socket (/var/run/lcf/lcfc.sock, a hostPath), and LCFC fans KV blocks out to the LCFT storage tier over the RDMA/TCP fabric.

fail_open cuts both ways

The connector is fail_open — if the fabric is unavailable, vLLM falls back to normal recompute, so serving is never blocked. The flip side is that a broken connector looks exactly like a healthy engine: "loaded" is not "connected". Always confirm with a warm-turn cache hit.