Endpoint Picker manifests

Manifests for the Endpoint Picker use-cases. Apply in file order; the chapters set MF to this folder.

File	Creates
01-namespace.yaml	namespace llm-d-pd-disaggregation
10-vllm.yaml	3 × vLLM Qwen/Qwen3-32B-FP8, one GPU each, label llm-d.ai/role: aggregated
20-epp-config.yaml	EndpointPickerConfig (prefix-cache-affinity-filter, token-load-scorer, active-request-scorer, max-score-picker)
21-epp-rbac.yaml	ServiceAccount, ClusterRole, ClusterRoleBinding for the EPP
22-epp.yaml	EPP Deployment pd-disaggregation-epp (llm-d-router-endpoint-picker:main)
23-epp-service.yaml	ClusterIP Service: 9002 ext_proc gRPC, 9003 health, 9090 metrics
30-inferencepool.yaml	InferencePool pd-disaggregation selecting the vLLM pods, endpointPickerRef → the EPP Service
50-gateway.yaml	kube-loxilb path only: GatewayClass loxilb-gc, Gateway inference-gw (listener :8080), HTTPRoute llm-route → the pool. kube-loxilb turns them into the rule
40-loxilb-igw.yaml	loxilb-igw DaemonSet in kube-system, control-plane nodes, hostNetwork, /etc/loxilb persisted
rule.sh	creates the test rule in EPP / CHWBL / RR mode from live pod IPs and the EPP ClusterIP

export VIP=192.0.2.10 PORT=8080 LB=http://$VIP:11111/netlox/v1   # VIP = the gateway node's IP
bash rule.sh epp      # or chwbl | rr | delete | show

Or, with kube-loxilb running --gatewayAPI --inferenceExtension, skip rule.sh and kubectl apply -f 50-gateway.yaml; delete the direct rule first if one exists on the same VIP:port.