cachebackend-lmcache.yaml

# SPDX-FileCopyrightText: 2026 The inference-cache Authors
#
# SPDX-License-Identifier: Apache-2.0

# Canonical vLLM PodLocal LMCache MP backend with a controller-managed Redis
# remote L3. CacheBackend injects one native LMCache MP server sidecar per
# matching engine Pod and points it at Redis; it never replaces the
# inference-owner's engine image. This sample explicitly selects Redis to keep
# cross-Pod sharing. It is not an automatic mapping from the legacy
# removed LMCacheServer provider, whose IP data plane had different semantics.
apiVersion: inferencecache.io/v1alpha1
kind: CacheBackend
metadata:
  labels:
    app.kubernetes.io/name: inference-cache
  name: cachebackend-lmcache
spec:
  runtime: VLLM
  type: LMCache
  integration:
    role: ReadWrite
  engineSelector:
    matchLabels:
      inferencecache.io/cache-domain: vllm-lmcache
  lmCache:
    topology: PodLocal
    chunkSizeTokens: 256
    podLocal:
      server:
        image: docker.io/lmcache/standalone@sha256:b813bf0bb616d1012b6a6edcbd4a44f1576dbbdaa857962e56d48b9f7c127d13
        port: 5555
        l1Capacity: 4Gi
        maxWorkers: 4
        resources:
          requests:
            cpu: "1"
            memory: 5Gi
          limits:
            cpu: "2"
            memory: 6Gi
  observation:
    modelID: meta-llama/Meta-Llama-3-8B-Instruct
  remoteStorage:
    provider: Redis
    ownership: Managed
    # Optional scheduling/security applies only to the controller-managed
    # provider Pod. It never mutates the inference-engine Pod.
    workload:
      nodeSelector:
        kubernetes.io/os: linux
    redis:
      image: docker.io/library/redis@sha256:e7723ff73d963f5cc6d9c4643ea3d989527a402a319239054e9472a7fb9219a2
      resources:
        requests:
          cpu: 500m
          memory: 4Gi
        limits:
          cpu: "2"
          memory: 8Gi