Add support for plugin configuration in the InferencePool helm chart

cmd/epp/runner/runner.go

@@ -348,6 +348,8 @@ func (r *Runner) parsePluginsConfiguration(ctx context.Context) error {
 		return fmt.Errorf("failed to load the configuration - %w", err)
 	}
 
+	setupLog.Info("Configuration file loaded", "config", config)
+
 	r.schedulerConfig, err = loader.LoadSchedulerConfig(config.SchedulingProfiles, handle)
 	if err != nil {
 		return fmt.Errorf("failed to create Scheduler configuration - %w", err)

config/charts/inferencepool/templates/epp-config.yaml

@@ -0,0 +1,85 @@
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: {{ include "gateway-api-inference-extension.name" . }}
+  namespace: {{ .Release.Namespace }}
+data:
+  default-plugins.yaml: |
+    apiVersion: inference.networking.x-k8s.io/v1alpha1
+    kind: EndpointPickerConfig
+    plugins:
+    - type: low-queue-filter
+      parameters:
+        threshold: 128
+    - type: lora-affinity-filter
+      parameters:
+        threshold: 0.999
+    - type: least-queue-filter
+    - type: least-kv-cache-filter
+    - type: decision-tree-filter
+      name: low-latency-filter
+      parameters:
+        current:
+          pluginRef: low-queue-filter
+        nextOnSuccess:
+          decisionTree:
+            current:
+              pluginRef: lora-affinity-filter
+            nextOnSuccessOrFailure:
+              decisionTree:
+                current:
+                  pluginRef: least-queue-filter
+                nextOnSuccessOrFailure:
+                  decisionTree:
+                    current:
+                      pluginRef: least-kv-cache-filter
+        nextOnFailure:
+          decisionTree:
+            current:
+              pluginRef: least-queue-filter
+            nextOnSuccessOrFailure:
+              decisionTree:
+                current:
+                  pluginRef: lora-affinity-filter
+                nextOnSuccessOrFailure:
+                  decisionTree:
+                    current:
+                      pluginRef: least-kv-cache-filter
+    - type: random-picker
+      parameters:
+        maxNumOfEndpoints: 1
+    - type: single-profile-handler
+    schedulingProfiles:
+    - name: default
+      plugins:
+      - pluginRef: low-latency-filter
+      - pluginRef: random-picker
+  plugins-v2.yaml: |
+    apiVersion: inference.networking.x-k8s.io/v1alpha1
+    kind: EndpointPickerConfig
+    plugins:
+    - type: queue-scorer
+    - type: kv-cache-scorer
+    - type: prefix-cache-scorer
+      parameters:
+        hashBlockSize: 64
+        maxPrefixBlocksToMatch: 256
+        lruCapacityPerServer: 31250
+    - type: max-score-picker
+      parameters:
+        maxNumOfEndpoints: 1
+    - type: single-profile-handler
+    schedulingProfiles:
+    - name: default
+      plugins:
+      - pluginRef: queue-scorer
+        weight: 1
+      - pluginRef: kv-cache-scorer
+        weight: 1
+      - pluginRef: prefix-cache-scorer
+        weight: 1
+      - pluginRef: max-score-picker
+  {{- if (hasKey .Values.inferenceExtension "pluginsCustomConfig") }}
+  {{- .Values.inferenceExtension.pluginsCustomConfig | toYaml | nindent 2 }}
+  {{- end }}
+

config/charts/inferencepool/templates/epp-deployment.yaml

@@ -35,6 +35,8 @@ spec:
         - "9003"
         - -metricsPort
         - "9090"
+        - -configFile
+        - "config/{{ .Values.inferenceExtension.pluginsConfigFile }}"
         # https://pkg.go.dev/flag#hdr-Command_line_flag_syntax; space is only for non-bool flags
         - "-enablePprof={{ .Values.inferenceExtension.enablePprof }}"
         {{- if eq (.Values.inferencePool.modelServerType | default "vllm") "triton-tensorrt-llm" }}
@@ -69,3 +71,10 @@ spec:
         - name: {{ $key }}
           value: {{ $value | quote }}
         {{- end }}
+        volumeMounts:
+        - name: plugins-config-volume
+          mountPath: "/config"
+      volumes:
+      - name: plugins-config-volume
+        configMap:
+          name: {{ include "gateway-api-inference-extension.name" . }}

config/charts/inferencepool/values.yaml

@@ -8,6 +8,26 @@ inferenceExtension:
   extProcPort: 9002
   env: {}
   enablePprof: true # Enable pprof handlers for profiling and debugging
+  # This is the plugins configuration file. 
+  pluginsConfigFile: "default-plugins.yaml"
+  # pluginsCustomConfig:
+  #   custom-plugins.yaml: |
+  #     apiVersion: inference.networking.x-k8s.io/v1alpha1
+  #     kind: EndpointPickerConfig
+  #     plugins:
+  #     - type: custom-scorer
+  #       parameters:
+  #         custom-threshold: 64
+  #     - type: max-score-picker
+  #     - type: single-profile-handler
+  #     schedulingProfiles:
+  #     - name: default
+  #       plugins:
+  #       - pluginRef: custom-scorer
+  #         weight: 1
+  #       - pluginRef: max-score-picker
+  #         weight: 1
+
   # Example environment variables:
   # env:
   #   KV_CACHE_SCORE_WEIGHT: "1"