cluster.yaml

# kind cluster for the vLLM + LMCache reference substrate.
#
# kind nodes are containers on the host's Docker daemon. kind does NOT expose
# host GPUs to node containers out of the box, so the GPU path below has two
# documented options. The CPU-only path (../manifests/cpu-local) needs no GPU.
#
#   kind create cluster --name inference-cache-substrate --config cluster.yaml
#
apiVersion: kind.x-k8s.io/v1alpha4
kind: Cluster
name: inference-cache-substrate
nodes:
  - role: control-plane
    # Expose the vLLM OpenAI API (30080) and the ZMQ KV-event PUB port (30557)
    # on localhost so the host can curl the API and subscribe to events without
    # `kubectl port-forward`. These map to the NodePort Service in
    # ../manifests/service.yaml.
    extraPortMappings:
      - containerPort: 30080   # vLLM OpenAI API
        hostPort: 30080
        protocol: TCP
      - containerPort: 30557   # vLLM KV-event ZMQ PUB
        hostPort: 30557
        protocol: TCP
# -----------------------------------------------------------------------------
# GPU on kind (for the real LMCache + vLLM run on an NVIDIA GPU host).
# kind has no first-class GPU flag; the supported approaches are:
#
#   (A) NVIDIA k8s-device-plugin on a GPU host. Install the NVIDIA Container
#       Toolkit on the host, set `nvidia` as the default Docker runtime, create
#       the kind cluster, then `helm install` the device plugin so pods can
#       request `nvidia.com/gpu`. See ../README.md "GPU on kind".
#
#   (B) Skip kind on the GPU box and run the same Pod spec on a managed cluster
#       (OKE) that already advertises `nvidia.com/gpu`. The manifests in
#       ../manifests are identical; only the cluster differs.
#
# This file is the CPU-friendly local-dev cluster. GPU node wiring is host- and
# environment-specific and lives in the runbook, not here.
# -----------------------------------------------------------------------------