lmcache--lmcache
wehub-resource-sync
127 行
4.7 KiB
YAML
127 行
4.7 KiB
YAML
apiVersion: lmcache.lmcache.ai/v1alpha1
|
|
kind: LMCacheEngine
|
|
metadata:
|
|
name: my-cache
|
|
namespace: default
|
|
spec:
|
|
# -- Container image --
|
|
# image:
|
|
# repository: lmcache/vllm-openai # default
|
|
# tag: latest # default
|
|
# pullPolicy: IfNotPresent # Always | Never | IfNotPresent
|
|
# imagePullSecrets:
|
|
# - name: my-registry-secret
|
|
|
|
# -- Server config --
|
|
# server:
|
|
# port: 5555 # default, ZMQ server port [1024, 65535]
|
|
# httpPort: 8080 # default, HTTP frontend port (health, admin)
|
|
# chunkSize: 256 # default, token chunk size
|
|
# maxWorkers: 1 # default, worker threads
|
|
# hashAlgorithm: blake3 # builtin | sha256_cbor | blake3
|
|
|
|
# -- L1 cache (REQUIRED) --
|
|
l1:
|
|
sizeGB: 60
|
|
|
|
# -- Eviction --
|
|
# eviction:
|
|
# policy: LRU # default (only supported value)
|
|
# triggerWatermark: 0.8 # default, range (0.0, 1.0]
|
|
# evictionRatio: 0.2 # default, range (0.0, 1.0]
|
|
|
|
# -- Prometheus monitoring --
|
|
# prometheus:
|
|
# enabled: true # default
|
|
# port: 9090 # default
|
|
# serviceMonitor:
|
|
# enabled: false # default, requires Prometheus Operator
|
|
# interval: 30s # default
|
|
# labels:
|
|
# release: kube-prometheus-stack
|
|
|
|
# -- L2 storage backend (single adapter) --
|
|
# l2Backend:
|
|
# # Option A: Redis/Valkey (native RESP connector)
|
|
# resp:
|
|
# host: redis.default.svc.cluster.local
|
|
# port: 6379
|
|
# numWorkers: 8 # default, C++ I/O threads
|
|
# maxCapacityGB: 0 # default (0 = no eviction tracking)
|
|
# authSecretRef: # optional, Secret with "username"/"password" keys
|
|
# name: redis-auth
|
|
# namespace: redis # omit if same namespace as LMCacheEngine
|
|
#
|
|
# # Option B: Raw escape hatch (for nixl_store, fs, mock, etc.)
|
|
# # raw:
|
|
# # type: nixl_store
|
|
# # config:
|
|
# # backend: "POSIX"
|
|
# # backend_params:
|
|
# # file_path: "/data/lmcache/l2"
|
|
# # use_direct_io: "false"
|
|
# # pool_size: 64
|
|
#
|
|
# # L2 policies
|
|
# storePolicy: default # default | skip_l1 (buffer-only mode)
|
|
# prefetchPolicy: default # default
|
|
# prefetchMaxInFlight: 8 # default, max concurrent L2->L1 loads
|
|
|
|
# -- Connection-injection webhook defaults (optional) --
|
|
# The mutating webhook always wires opted-in vLLM pods to this engine
|
|
# (--kv-transfer-config, hostIPC, PYTHONHASHSEED). Set injection.payloadImage to
|
|
# ALSO stage an lmcache code tree into the vLLM container (emptyDir + init
|
|
# container + read-only mount + PYTHONPATH), pinning vLLM's lmcache to a
|
|
# specific build so client and server match. Leave unset for connection-only.
|
|
# injection:
|
|
# payloadImage: # SEPARATE image: ships lmcache under /payload,
|
|
# repository: registry.example.com/lmcache/lmcache-payload # REQUIRED (no default)
|
|
# tag: pinned # default: latest
|
|
# pullPolicy: IfNotPresent # Always | Never | IfNotPresent
|
|
# imagePullSecrets: # for a private payload image
|
|
# - name: lmcache-payload-pull
|
|
# targetContainer: vllm # default: first container
|
|
|
|
# -- Resource overrides (auto-computed from l1.sizeGB if omitted) --
|
|
# resourceOverrides:
|
|
# requests:
|
|
# memory: 70Gi
|
|
# cpu: "4"
|
|
# limits:
|
|
# memory: 100Gi
|
|
|
|
# -- Logging --
|
|
# logLevel: INFO # DEBUG | INFO | WARNING | ERROR
|
|
|
|
# -- Scheduling --
|
|
# nodeSelector: # defaults to nvidia.com/gpu.present: "true"
|
|
# nvidia.com/gpu.present: "true"
|
|
# affinity: {}
|
|
# tolerations: []
|
|
|
|
# -- Pod overrides --
|
|
# WARNING: Do NOT mount an emptyDir (or any volume) at /dev/shm. The operator
|
|
# sets hostIPC: true so the container shares the host's /dev/shm. Mounting an
|
|
# emptyDir would shadow the host's /dev/shm with a private tmpfs, breaking
|
|
# CUDA IPC between LMCache and vLLM pods (cudaIpcOpenMemHandle fails because
|
|
# IPC handles from one pod become invisible to the other). The same applies
|
|
# to vLLM pods that connect to this LMCache instance.
|
|
# env: []
|
|
# volumes: []
|
|
# volumeMounts: []
|
|
# podAnnotations: {}
|
|
# podLabels: {}
|
|
# serviceAccountName: ""
|
|
# priorityClassName: ""
|
|
|
|
# -- Security --
|
|
# Run the engine container in privileged mode. Defaults to false. On most
|
|
# clusters runtimeClassName: nvidia + NVIDIA_VISIBLE_DEVICES=all is enough for
|
|
# the engine to see all GPUs; set this to true only on clusters where it
|
|
# cannot see them otherwise. Requires the namespace to allow the privileged
|
|
# Pod Security Standard.
|
|
# privileged: false
|
|
|
|
# -- Extra CLI flags (appended last, can override any auto-generated flag) --
|
|
# extraArgs: []
|