wehub-resource-sync
51 lines
2.0 KiB
YAML
51 lines
2.0 KiB
YAML
# CacheBlendEngine used by the e2e_gpu CacheBlend integration spec.
|
|
#
|
|
# NewCBEFromFixture decodes this YAML into a typed CacheBlendEngine and
|
|
# overrides metadata.name / metadata.namespace. The spec then overrides the
|
|
# two env-driven, deployment-specific bits before applying:
|
|
#
|
|
# - spec.image (CACHEBLEND_ENGINE_IMAGE; the lmcache
|
|
# blend server image — public by default)
|
|
# - spec.injection.payloadImage (CACHEBLEND_PAYLOAD_IMAGE; the PRIVATE
|
|
# cacheblend-plugin init-container image)
|
|
#
|
|
# spec.injection.imagePullSecrets references the docker-registry Secret the
|
|
# spec creates in this namespace from CACHEBLEND_REGISTRY_USER/TOKEN. The
|
|
# mutating webhook appends it to the vLLM pod so the payload init container
|
|
# can pull the private image.
|
|
apiVersion: lmcache.lmcache.ai/v1alpha1
|
|
kind: CacheBlendEngine
|
|
metadata:
|
|
name: cb-integration
|
|
spec:
|
|
# The lmcache blend server image. Public; pinned to the same image the
|
|
# vLLM workload uses so a fresh GPU node pulls it only once. Nightly to stay
|
|
# in the CacheBlend compatibility window with the latest-nightly payload.
|
|
image:
|
|
repository: lmcache/vllm-openai
|
|
tag: latest-nightly
|
|
pullPolicy: IfNotPresent
|
|
|
|
# L1 cache is REQUIRED. 8 GiB is plenty for the small smoke model.
|
|
l1:
|
|
sizeGB: 8
|
|
|
|
injection:
|
|
# The PRIVATE cacheblend-plugin init-container image. The spec overrides
|
|
# repository/tag from CACHEBLEND_PAYLOAD_IMAGE; these are the defaults.
|
|
payloadImage:
|
|
repository: tensormesh/cacheblend-plugin
|
|
tag: latest-nightly
|
|
pullPolicy: IfNotPresent
|
|
# Pull secret for the private payload image. Created by the spec in this
|
|
# namespace (name must match the spec's pullSecretName constant).
|
|
imagePullSecrets:
|
|
- name: cacheblend-registry
|
|
|
|
# Place the engine on the GPU worker so the vLLM pod (same nodeSelector)
|
|
# co-schedules — CacheBlend's CUDA-IPC KV handoff requires colocation.
|
|
nodeSelector:
|
|
nvidia.com/gpu.present: "true"
|
|
|
|
logLevel: INFO
|