feat(coordinator): serve encode/prefill/decode from a single EPP

The coordinator's per-phase model required three separate EPP
instances (one each for encode, prefill, decode), each with its own
Deployment, Service, and InferencePool, even though every
coordinator-to-worker call already carries an EPP-Phase header naming
the phase. Combining them into one EPP simplifies deployment and is a
prerequisite for p2p KV transfer (non-deferred decode).

Add a header-phase-profile-handler EPP plugin that runs the single
scheduling profile named by the EPP-Phase header, replacing the
per-instance single-profile-handler. Merge the three EPPs' scheduling
configs into one EndpointPickerConfig with encode/prefill/decode
profiles, each still gated by its existing by-label role filter so it
only sees its own role's pods out of the combined pool. Replace the
three epd-pools/{encode,prefill,decode} deploy components with one
epd-pool component: a single EPP Deployment/Service and one
InferencePool selecting the app label the worker Deployments already
share. Collapse the e2e Envoy test routing from three EPP-Phase-matched
ext_proc clusters to one, and update the coordinator e2e test
setup/assertions and architecture docs to match.

Signed-off-by: roytman <roytman@il.ibm.com>

Author: roytman

cmd/epp/runner/runner.go

@@ -115,6 +115,7 @@ import (
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/picker/weightedrandom"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/profilehandler/dataparallel"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/profilehandler/disagg"
+	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/profilehandler/headerphase"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/profilehandler/single"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/scorer/activerequest"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/scorer/contextlengthaware"
@@ -579,6 +580,7 @@ func (r *Runner) registerInTreePlugins() {
 	fwkplugin.Register(random.RandomPickerType, random.RandomPickerFactory)
 	fwkplugin.Register(weightedrandom.WeightedRandomPickerType, weightedrandom.WeightedRandomPickerFactory)
 	fwkplugin.Register(single.SingleProfileHandlerType, single.SingleProfileHandlerFactory)
+	fwkplugin.Register(headerphase.HeaderPhaseProfileHandlerType, headerphase.Factory)
 	fwkplugin.Register(disagg.DisaggHeadersHandlerType, disagg.HeadersHandlerFactory)                     //nolint:staticcheck // intentional: keep backward compatibility
 	fwkplugin.Register(disagg.PrefillHeaderHandlerType, disagg.HeadersHandlerFactory)                     //nolint:staticcheck // intentional: keep backward compatibility
 	fwkplugin.RegisterWithPluginDependencies(disagg.PdProfileHandlerType, disagg.PdProfileHandlerFactory, //nolint:staticcheck // intentional: keep backward compatibility

…rence-gateway/epd-pools/prefill/epp.yaml …ents/inference-gateway/epd-pool/epp.yamldeploy/coordinator/components/inference-gateway/epd-pools/prefill/epp.yaml renamed to deploy/coordinator/components/inference-gateway/epd-pool/epp.yaml deploy/coordinator/components/inference-gateway/epd-pools/prefill/epp.yaml renamed to deploy/coordinator/components/inference-gateway/epd-pool/epp.yaml

@@ -1,15 +1,15 @@
 apiVersion: v1
 kind: ServiceAccount
 metadata:
-  name: ${EPP_NAME}-prefill
+  name: ${EPP_NAME}
 ---
 apiVersion: rbac.authorization.k8s.io/v1
 kind: RoleBinding
 metadata:
-  name: ${EPP_NAME}-prefill-binding
+  name: ${EPP_NAME}-binding
 subjects:
 - kind: ServiceAccount
-  name: ${EPP_NAME}-prefill
+  name: ${EPP_NAME}
 roleRef:
   apiGroup: rbac.authorization.k8s.io
   kind: Role
@@ -18,12 +18,12 @@ roleRef:
 apiVersion: v1
 kind: Service
 metadata:
-  name: ${EPP_NAME}-prefill
+  name: ${EPP_NAME}
   labels:
-    app: ${EPP_NAME}-prefill
+    app: ${EPP_NAME}
 spec:
   selector:
-    app: ${EPP_NAME}-prefill
+    app: ${EPP_NAME}
   ports:
   - name: default
     protocol: TCP
@@ -44,28 +44,28 @@ spec:
 apiVersion: apps/v1
 kind: Deployment
 metadata:
-  name: ${EPP_NAME}-prefill
+  name: ${EPP_NAME}
   labels:
-    app: ${EPP_NAME}-prefill
+    app: ${EPP_NAME}
 spec:
   replicas: 1
   selector:
     matchLabels:
-      app: ${EPP_NAME}-prefill
+      app: ${EPP_NAME}
   template:
     metadata:
       labels:
-        app: ${EPP_NAME}-prefill
+        app: ${EPP_NAME}
     spec:
-      serviceAccountName: ${EPP_NAME}-prefill
+      serviceAccountName: ${EPP_NAME}
       terminationGracePeriodSeconds: 130
       containers:
       - name: epp
         image: ${EPP_IMAGE}
         imagePullPolicy: IfNotPresent
         args:
         - --pool-name
-        - "${POOL_NAME}-prefill"
+        - "${POOL_NAME}"
         - --pool-namespace
         - "${NAMESPACE}"
         - --v
@@ -103,4 +103,4 @@ spec:
       volumes:
       - name: epp-config
         configMap:
-          name: epp-config-prefill
+          name: epp-config

…way/epd-pools/decode/inference-pool.yaml …nce-gateway/epd-pool/inference-pool.yamldeploy/coordinator/components/inference-gateway/epd-pools/decode/inference-pool.yaml renamed to deploy/coordinator/components/inference-gateway/epd-pool/inference-pool.yaml deploy/coordinator/components/inference-gateway/epd-pools/decode/inference-pool.yaml renamed to deploy/coordinator/components/inference-gateway/epd-pool/inference-pool.yaml

@@ -1,13 +1,13 @@
 apiVersion: inference.networking.k8s.io/v1
 kind: InferencePool
 metadata:
-  name: ${POOL_NAME}-decode
+  name: ${POOL_NAME}
 spec:
   selector:
     matchLabels:
-      llm-d.ai/role: decode
+      app: ${POOL_NAME}
   endpointPickerRef:
-    name: ${EPP_NAME}-decode
+    name: ${EPP_NAME}
     kind: Service
     port:
       number: 9002