change model name in e2e tests to be real small model supporting multi-modal requests

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

Author: mayabar

scripts/kind-dev-env.sh

@@ -37,14 +37,7 @@ EPP_IMAGE="${EPP_IMAGE:-${IMAGE_REGISTRY}/llm-d-router-endpoint-picker:${EPP_TAG
 export EPP_IMAGE
 
 # Set the model name to deploy.
-# When Encode disaggregation is enabled (multimodal pipeline), default to a
-# multimodal model. Otherwise use the standard text-only model.
-# Note: DISAGG_E/DISAGG_P are set later in this script, so read the raw env vars here.
-if [ "${DISAGG_E:-false}" == "true" ] || [ "${EPD_ENABLED:-false}" == "true" ] || [ "${EPD_ENABLED:-false}" == "\"true\"" ]; then
-  export MODEL_NAME="${MODEL_NAME:-Qwen/Qwen3-VL-2B-Instruct}"
-else
-  export MODEL_NAME="${MODEL_NAME:-TinyLlama/TinyLlama-1.1B-Chat-v1.0}"
-fi
+export MODEL_NAME="${MODEL_NAME:-Qwen/Qwen3-VL-2B-Instruct}"
 # Extract model family (e.g., "meta-llama" from "meta-llama/Llama-3.1-8B-Instruct")
 export MODEL_FAMILY="${MODEL_NAME%%/*}"
 # Extract model ID (e.g., "Llama-3.1-8B-Instruct")

test/e2e/configs_test.go

@@ -1,5 +1,9 @@
 package e2e
 
+import (
+	"github.com/llm-d/llm-d-router/test/utils"
+)
+
 // Simple EPP configuration for running without P/D
 const simpleConfig = `apiVersion: llm-d.ai/v1alpha1
 kind: EndpointPickerConfig
@@ -185,7 +189,7 @@ kind: EndpointPickerConfig
 plugins:
 - type: token-producer
   parameters:
-    modelName: Qwen/Qwen2.5-1.5B-Instruct
+    modelName: ` + utils.ModelName + `
     vllm:
       url: http://localhost:8000
 - type: precise-prefix-cache-scorer
@@ -217,7 +221,7 @@ kind: EndpointPickerConfig
 plugins:
 - type: token-producer
   parameters:
-    modelName: Qwen/Qwen2.5-1.5B-Instruct
+    modelName: ` + utils.ModelName + `
     vllm:
       url: http://localhost:8000
 - type: precise-prefix-cache-scorer

test/e2e/disruption_test.go

@@ -33,7 +33,7 @@ var disruptionClient = &http.Client{Timeout: 10 * time.Second}
 
 // sendRawCompletion sends a completion request and returns the HTTP status code.
 func sendRawCompletion() (int, error) {
-	body := fmt.Sprintf(`{"model":"%s","prompt":"%s","max_tokens":10}`, simModelName, simplePrompt)
+	body := fmt.Sprintf(`{"model":"%s","prompt":"%s","max_tokens":10}`, testutils.ModelName, simplePrompt)
 	req, err := http.NewRequest(http.MethodPost, fmt.Sprintf("http://localhost:%s/v1/completions", port), strings.NewReader(body))
 	if err != nil {
 		return 0, err
@@ -99,7 +99,7 @@ var _ = ginkgo.Describe("Disruption tests", ginkgo.Ordered, ginkgo.Label("Disrup
 			gomega.Expect(decodePods).Should(gomega.HaveLen(2))
 
 			ginkgo.By("Verifying requests route successfully before disruption")
-			nsHdr, _, _ := runCompletion(simplePrompt, simModelName)
+			nsHdr, _, _ := runCompletion(simplePrompt, testutils.ModelName)
 			gomega.Expect(nsHdr).Should(gomega.Equal(nsName))
 
 			targetPod := decodePods[0]
@@ -112,7 +112,7 @@ var _ = ginkgo.Describe("Disruption tests", ginkgo.Ordered, ginkgo.Label("Disrup
 
 			ginkgo.By("Verifying new requests eventually route to a pod other than the killed one")
 			gomega.Eventually(func() error {
-				nsHdr, podHdr, _, err := tryCompletion(simplePrompt, simModelName)
+				nsHdr, podHdr, _, err := tryCompletion(simplePrompt, testutils.ModelName)
 				if err != nil {
 					return err
 				}
@@ -133,7 +133,7 @@ var _ = ginkgo.Describe("Disruption tests", ginkgo.Ordered, ginkgo.Label("Disrup
 
 			ginkgo.By("Verifying requests succeed after recovery")
 			for range 3 {
-				nsHdr, _, _ = runCompletion(simplePrompt, simModelName)
+				nsHdr, _, _ = runCompletion(simplePrompt, testutils.ModelName)
 				gomega.Expect(nsHdr).Should(gomega.Equal(nsName))
 			}
 		})
@@ -185,7 +185,7 @@ var _ = ginkgo.Describe("Disruption tests", ginkgo.Ordered, ginkgo.Label("Disrup
 
 			ginkgo.By("Verifying requests succeed after recovery")
 			for range 3 {
-				nsHdr, _, _ := runCompletion(simplePrompt, simModelName)
+				nsHdr, _, _ := runCompletion(simplePrompt, testutils.ModelName)
 				gomega.Expect(nsHdr).Should(gomega.Equal(nsName))
 			}
 		})
@@ -205,7 +205,7 @@ var _ = ginkgo.Describe("Disruption tests", ginkgo.Ordered, ginkgo.Label("Disrup
 			gomega.Expect(decodePods).Should(gomega.HaveLen(1))
 
 			ginkgo.By("Verifying requests succeed before disruption")
-			nsHdr, _, _ := runCompletion(simplePrompt, simModelName)
+			nsHdr, _, _ := runCompletion(simplePrompt, testutils.ModelName)
 			gomega.Expect(nsHdr).Should(gomega.Equal(nsName))
 
 			ginkgo.By("Scaling deployment to zero")
@@ -231,7 +231,7 @@ var _ = ginkgo.Describe("Disruption tests", ginkgo.Ordered, ginkgo.Label("Disrup
 
 			ginkgo.By("Verifying requests succeed after recovery")
 			gomega.Eventually(func() string {
-				nsHdr, _, _ := runCompletion(simplePrompt, simModelName)
+				nsHdr, _, _ := runCompletion(simplePrompt, testutils.ModelName)
 				return nsHdr
 			}, eppRecoveryTimeout, 2*time.Second).Should(gomega.Equal(nsName))
 		})
@@ -248,7 +248,7 @@ var _ = ginkgo.Describe("Disruption tests", ginkgo.Ordered, ginkgo.Label("Disrup
 			ginkgo.DeferCleanup(testutils.DeleteObjects, testConfig, epp)
 
 			ginkgo.By("Verifying requests succeed before EPP disruption")
-			nsHdr, _, _ := runCompletion(simplePrompt, simModelName)
+			nsHdr, _, _ := runCompletion(simplePrompt, testutils.ModelName)
 			gomega.Expect(nsHdr).Should(gomega.Equal(nsName))
 
 			ginkgo.By("Finding EPP pod")
@@ -297,7 +297,7 @@ var _ = ginkgo.Describe("Disruption tests", ginkgo.Ordered, ginkgo.Label("Disrup
 			ginkgo.DeferCleanup(testutils.DeleteObjects, testConfig, epp)
 
 			ginkgo.By("Verifying requests succeed before disruption")
-			nsHdr, _, _ := runCompletion(simplePrompt, simModelName)
+			nsHdr, _, _ := runCompletion(simplePrompt, testutils.ModelName)
 			gomega.Expect(nsHdr).Should(gomega.Equal(nsName))
 
 			ginkgo.By("Starting background traffic")
@@ -341,7 +341,7 @@ var _ = ginkgo.Describe("Disruption tests", ginkgo.Ordered, ginkgo.Label("Disrup
 
 func sendStreamingCompletion(connected chan<- string) error {
 	longPrompt := strings.Repeat("This is a longer prompt to keep the stream open. ", 20)
-	body := fmt.Sprintf(`{"model":"%s","prompt":"%s","max_tokens":100,"stream":true}`, simModelName, longPrompt)
+	body := fmt.Sprintf(`{"model":"%s","prompt":"%s","max_tokens":100,"stream":true}`, testutils.ModelName, longPrompt)
 	req, err := http.NewRequest(http.MethodPost, fmt.Sprintf("http://localhost:%s/v1/completions", port), strings.NewReader(body))
 	if err != nil {
 		connected <- ""

test/e2e/e2e_suite_test.go

@@ -36,10 +36,6 @@ const (
 	crdKustomizePath = "../../config/crd"
 	// inferExtManifest is the manifest for the inference extension test resources.
 	inferExtManifest = "../../deploy/components/inference-gateway/inference-pools.yaml"
-	// simModelName is the test model name.
-	simModelName = "food-review"
-	// kvModelName is the model name used in KV tests.
-	kvModelName = "Qwen/Qwen2.5-1.5B-Instruct"
 	// envoyManifest is the manifest for the envoy proxy test resources.
 	envoyManifest = "../../deploy/environments/dev/e2e-infra/envoy.yaml"
 	// eppManifest is the manifest for the deployment of the EPP
@@ -303,7 +299,7 @@ func createEnvoy() {
 }
 
 func createInferencePool(numTargetPorts int, toDelete bool) []string {
-	poolName := simModelName + "-inference-pool"
+	poolName := testutils.ModelServerName + "-inference-pool"
 
 	if toDelete {
 		objName := []string{"inferencepool/" + poolName}