review

Signed-off-by: Maya Barnea <mayab@il.ibm.com>

Author: mayabar

test/e2e/configs_test.go

@@ -1,5 +1,7 @@
 package e2e
 
+import "fmt"
+
 // Simple EPP configuration for running without P/D
 const simpleConfig = `apiVersion: llm-d.ai/v1alpha1
 kind: EndpointPickerConfig
@@ -179,15 +181,17 @@ schedulingProfiles:
     weight: 2
 `
 
-// EPP config for running with precise prefix scoring (i.e. KV events).
-const kvConfig = `apiVersion: llm-d.ai/v1alpha1
+// kvConfig returns the EPP config for running with precise prefix scoring (i.e. KV events).
+// The render URL is built from vllmRenderPort so VLLM_RENDER_PORT is respected.
+func kvConfig() string {
+	return fmt.Sprintf(`apiVersion: llm-d.ai/v1alpha1
 kind: EndpointPickerConfig
 plugins:
 - type: token-producer
   parameters:
     modelName: Qwen/Qwen2.5-1.5B-Instruct
     vllm:
-      url: http://vllm-render:8082
+      url: http://vllm-render:%s
 - type: precise-prefix-cache-scorer
   parameters:
     tokenProcessorConfig:
@@ -209,17 +213,20 @@ schedulingProfiles:
   - pluginRef: max-score-picker
   - pluginRef: precise-prefix-cache-scorer
     weight: 10
-`
+`, vllmRenderPort)
+}
 
-// Alias of kvConfig retained for tests that reference the external-tokenizer name.
-const kvExternalTokenizerConfig = `apiVersion: llm-d.ai/v1alpha1
+// kvExternalTokenizerConfig returns the EPP config for the external-tokenizer DataProducer variant.
+// The render URL is built from vllmRenderPort so VLLM_RENDER_PORT is respected.
+func kvExternalTokenizerConfig() string {
+	return fmt.Sprintf(`apiVersion: llm-d.ai/v1alpha1
 kind: EndpointPickerConfig
 plugins:
 - type: token-producer
   parameters:
     modelName: Qwen/Qwen2.5-1.5B-Instruct
     vllm:
-      url: http://vllm-render:8082
+      url: http://vllm-render:%s
 - type: precise-prefix-cache-scorer
   parameters:
     tokenProcessorConfig:
@@ -240,7 +247,8 @@ schedulingProfiles:
   - pluginRef: max-score-picker
   - pluginRef: precise-prefix-cache-scorer
     weight: 10
-`
+`, vllmRenderPort)
+}
 
 // EPP configuration for running scale model server test
 const scaleConfig = `apiVersion: llm-d.ai/v1alpha1

test/e2e/e2e_test.go

@@ -774,7 +774,7 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
 		ginkgo.It("should run successfully", func() {
 			infPoolObjects = createInferencePool(1, true)
 
-			epp := createEndPointPicker(kvConfig)
+			epp := createEndPointPicker(kvConfig())
 
 			modelServers := createModelServersDecodeKV(1)
 			time.Sleep(5 * time.Second) // wait for model server(s) to become ready
@@ -798,7 +798,7 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
 		ginkgo.It("should run successfully", func() {
 			infPoolObjects = createInferencePool(1, true)
 
-			epp := createEndPointPicker(kvExternalTokenizerConfig)
+			epp := createEndPointPicker(kvExternalTokenizerConfig())
 
 			modelServers := createModelServersDecodeKV(1)
 			time.Sleep(5 * time.Second) // wait for model server(s) to become ready

test/e2e/setup_test.go

@@ -34,6 +34,8 @@ func createModelServersFromKustomize(kustomizeDir string, extra map[string]strin
 		"${VLLM_EXTRA_ARGS_E}":       "--force-dummy-tokenizer",
 		"${VLLM_EXTRA_ARGS_P}":       "--force-dummy-tokenizer",
 		"${VLLM_EXTRA_ARGS_D}":       "--force-dummy-tokenizer",
+		"${VLLM_RENDER_URL}":         fmt.Sprintf("http://vllm-render.%s.svc.cluster.local:%s", nsName, vllmRenderPort),
+		"${VLLM_RENDER_PORT}":        vllmRenderPort,
 	}
 	for k, v := range extra {
 		subs[k] = v