wehub-resource-sync
95 lines
3.9 KiB
YAML
95 lines
3.9 KiB
YAML
apiVersion: lmcache.lmcache.ai/v1alpha1
|
|
kind: CacheBlendEngine
|
|
metadata:
|
|
name: my-cacheblend
|
|
namespace: default
|
|
spec:
|
|
# -- Container image (the LMCache blend_v3 ENGINE image) --
|
|
# The engine runs `lmcache server --engine-type blend`. If this image lives
|
|
# in a PRIVATE registry, set imagePullSecrets (the Secret must exist in this
|
|
# namespace).
|
|
image:
|
|
repository: lmcache/vllm-openai # default
|
|
tag: latest-nightly # default
|
|
pullPolicy: IfNotPresent # Always | Never | IfNotPresent
|
|
# imagePullSecrets:
|
|
# - name: my-registry-secret
|
|
|
|
# -- Server config --
|
|
# server:
|
|
# port: 6566 # MP port the CB connector dials (default 6566)
|
|
# httpPort: 8080 # HTTP frontend port (health, admin)
|
|
# chunkSize: 256 # FIXED at 256 (must equal vLLM --block-size * 4)
|
|
# maxWorkers: 1 # default, worker threads
|
|
# hashAlgorithm: blake3 # builtin | sha256_cbor | blake3
|
|
|
|
# -- L1 cache (REQUIRED) --
|
|
l1:
|
|
sizeGB: 60
|
|
|
|
# -- CacheBlend tunables (injected into the vLLM connector config) --
|
|
# blend:
|
|
# checkLayer: 1 # default, cb.check_layer
|
|
# recompRatio: 0.15 # default, cb.recomp_ratio
|
|
|
|
# -- Injection: how the webhook injects the plugin into annotated vLLM pods --
|
|
injection:
|
|
# payloadImage is the PRIVATE cacheblend-plugin init-container image (not
|
|
# open-source) — repository/tag/pullPolicy, like spec.image. Set repository
|
|
# to YOUR cacheblend-plugin image and pin the tag to the target vLLM/lmcache
|
|
# window. The webhook injects it as an init container into vLLM pods.
|
|
payloadImage:
|
|
repository: lmcache/cacheblend-plugin
|
|
tag: latest
|
|
pullPolicy: IfNotPresent # Always | Never | IfNotPresent
|
|
# imagePullSecrets pulls the PRIVATE payload image. The webhook appends these
|
|
# to the vLLM pod's spec.imagePullSecrets — the Secret(s) must exist in the
|
|
# vLLM pod's namespace.
|
|
# imagePullSecrets:
|
|
# - name: my-registry-secret
|
|
# targetContainer: "" # default: first container; name the vLLM container
|
|
# cudagraph: eager # eager | piecewise | full_decode_only (never full)
|
|
|
|
# -- 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
|
|
|
|
# -- Scheduling (place one engine per GPU node; vLLM pods must land on these nodes) --
|
|
# nodeSelector:
|
|
# nvidia.com/gpu.present: "true"
|
|
|
|
# -- Logging --
|
|
# logLevel: INFO # DEBUG | INFO | WARNING | ERROR
|
|
|
|
# -- Security --
|
|
# Run the engine container in privileged mode. Defaults to false. Set to true
|
|
# only on clusters where the engine cannot see the node GPUs otherwise
|
|
# (runtimeClassName: nvidia + NVIDIA_VISIBLE_DEVICES=all is enough on most).
|
|
# privileged: false
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# To enable CacheBlend on a vLLM pod, label it for the webhook and annotate it
|
|
# with this engine's name. The pod (and this engine) must run in a namespace
|
|
# labeled `pod-security.kubernetes.io/enforce=privileged` (hostIPC is injected).
|
|
#
|
|
# metadata:
|
|
# labels:
|
|
# lmcache.ai/cacheblend-inject: "true"
|
|
# annotations:
|
|
# lmcache.ai/cacheblend-engine: "my-cacheblend"
|
|
# # lmcache.ai/cacheblend-container: "vllm" # optional
|
|
# # lmcache.ai/cacheblend-image-pull-secrets: "..." # optional override
|
|
#
|
|
# The vLLM container must launch via the image ENTRYPOINT (args-only) — a
|
|
# `command: ["/bin/sh","-c", ...]` wrapper is skipped (injection cannot reach it).
|
|
# ---------------------------------------------------------------------------
|