values-reference.yaml

# Upstream "reference Helm chart" path (vLLM Production Stack).
#
#   helm repo add vllm https://vllm-project.github.io/production-stack
#   helm install vllm-substrate vllm/vllm-stack \
#     --version <pin>  -n cache-substrate --create-namespace \
#     -f values-reference.yaml
#
# This is the upstream-supported way to deploy vLLM+LMCache and is useful as an
# alternative to, or cross-check of, the raw manifests (../manifests). The chart
# also ships a built-in router, which this reference disables — the goal here is
# cache state and events, not routing.
#
# Requires a GPU node advertising nvidia.com/gpu and an `hf-token` secret with
# key `token`.
#
# NOTE: field names below (extraArgs, extraPorts, lmcacheConfig, vllmConfig)
# follow the Production-Stack chart schema as of the pinned version — verify
# against `helm show values vllm/vllm-stack --version <pin>` before relying on
# them; the chart's schema evolves. The hand-written ../manifests are the
# primary reference; this chart path is an optional alternative.
servingEngineSpec:
  runtimeClassName: ""        # set to "nvidia" if your cluster requires it
  modelSpec:
    - name: "llama8b"
      repository: "lmcache/vllm-openai"
      # PLACEHOLDER — replace with a pinned digest before any GPU deployment. Left
      # non-resolvable on purpose so the values file can't run an unpinned image.
      tag: "REPLACE_WITH_PINNED_DIGEST"   # e.g. set repository to ...@sha256:<digest> per the chart's pinning support; see ../VERSIONS.md
      modelURL: "meta-llama/Llama-3.1-8B-Instruct"
      replicaCount: 1
      requestGPU: 1
      vllmConfig:
        v1: 1                 # required for KV events + LMCacheConnectorV1
        enablePrefixCaching: true
        maxModelLen: 16384
        # The chart does not template --kv-events-config; inject it verbatim:
        extraArgs:
          - "--kv-events-config"
          - '{"enable_kv_cache_events":true,"publisher":"zmq","endpoint":"tcp://*:5557","replay_endpoint":"tcp://*:5558","buffer_steps":10000,"topic":"kv-events"}'
      lmcacheConfig:
        enabled: true
        cpuOffloadingBufferSize: "20"   # GiB
      hf_token:
        secretName: "hf-token"
        secretKey: "token"
      # Expose the ZMQ PUB port on the engine pod/Service (chart default exposes 8000 only).
      extraPorts:
        - name: kv-events
          containerPort: 5557
        - name: kv-replay
          containerPort: 5558

# We do NOT use the chart's router — the gateway decides routing; inference-cache
# only describes cache state.
routerSpec:
  enableRouter: false