feat: Add GAIE nightly integration test [DEP-423] (#7458)

Signed-off-by: Anna Tchernych <atchernych@nvidia.com>

.github/actions/dynamo-deploy-test/action.yml

@@ -31,17 +31,28 @@ inputs:
 
   framework:
     description: 'Framework name (vllm, sglang, trtllm)'
-    required: true
+    required: false
+    default: ''
   profile:
     description: 'Deployment profile (e.g., disagg_router, agg)'
-    required: true
+    required: false
+    default: ''
   image:
     description: 'Full container image reference for the framework runtime'
-    required: true
+    required: false
+    default: ''
   platform_arch:
     description: 'Platform architecture (amd64, arm64)'
     required: false
     default: 'amd64'
+  test_name:
+    description: 'Name for artifact naming. Defaults to {framework}_{profile}.'
+    required: false
+    default: ''
+  extra_pytest_args:
+    description: 'Additional pytest arguments (e.g., --frontend-image=...)'
+    required: false
+    default: ''
 
 runs:
   using: "composite"
@@ -82,16 +93,33 @@ runs:
         FRAMEWORK: ${{ inputs.framework }}
         PROFILE: ${{ inputs.profile }}
         IMAGE: ${{ inputs.image }}
+        TEST_NAME: ${{ inputs.test_name }}
+        EXTRA_PYTEST_ARGS: ${{ inputs.extra_pytest_args }}
       run: |
         mkdir -p test-results
+
+        PYTEST_ARGS=""
+        if [ -n "${FRAMEWORK}" ]; then
+          PYTEST_ARGS+=" --framework=${FRAMEWORK}"
+        fi
+        if [ -n "${PROFILE}" ]; then
+          PYTEST_ARGS+=" --profile=${PROFILE}"
+        fi
+        if [ -n "${IMAGE}" ]; then
+          PYTEST_ARGS+=" --image=${IMAGE}"
+        fi
+
+        if [ -z "${TEST_NAME}" ]; then
+          TEST_NAME="${FRAMEWORK:-unknown}_${PROFILE:-unknown}"
+        fi
+
         pytest tests/deploy/test_deploy.py \
-          --framework="${FRAMEWORK}" \
-          --profile="${PROFILE}" \
-          --image="${IMAGE}" \
           --namespace="${NAMESPACE}" \
+          ${PYTEST_ARGS} \
+          ${EXTRA_PYTEST_ARGS} \
           -v -s \
           --durations=10 \
-          --junitxml=test-results/pytest_deploy_${FRAMEWORK}_${PROFILE}_${{ inputs.platform_arch }}_${{ github.run_id }}_${{ job.check_run_id }}.xml \
+          --junitxml=test-results/pytest_deploy_${TEST_NAME}_${{ inputs.platform_arch }}_${{ github.run_id }}_${{ job.check_run_id }}.xml \
           --log-cli-level=INFO
 
     - name: Cleanup Deployment
@@ -124,10 +152,25 @@ runs:
         fi
         echo "::endgroup::"
 
+
+    - name: Determine test name for artifacts
+      id: test-name
+      if: always()
+      shell: bash
+      env:
+        TEST_NAME: ${{ inputs.test_name }}
+        FRAMEWORK: ${{ inputs.framework }}
+        PROFILE: ${{ inputs.profile }}
+      run: |
+        if [ -z "${TEST_NAME}" ]; then
+          TEST_NAME="${FRAMEWORK:-unknown}_${PROFILE:-unknown}"
+        fi
+        echo "name=${TEST_NAME}" >> $GITHUB_OUTPUT
+
     - name: Upload Test Results
       uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f #v6
       if: always()
       with:
-        name: test-results-${{ inputs.framework }}-${{ inputs.profile }}-${{ inputs.platform_arch }}-${{ github.run_id }}-${{ job.check_run_id }}
-        path: test-results/pytest_deploy_${{ inputs.framework }}_${{ inputs.profile }}_${{ inputs.platform_arch }}_${{ github.run_id }}_${{ job.check_run_id }}.xml
+        name: test-results-${{ steps.test-name.outputs.name }}-${{ inputs.platform_arch }}-${{ github.run_id }}-${{ job.check_run_id }}
+        path: test-results/pytest_deploy_${{ steps.test-name.outputs.name }}_${{ inputs.platform_arch }}_${{ github.run_id }}_${{ job.check_run_id }}.xml
         retention-days: 7

.github/workflows/post-merge-ci.yml

@@ -190,6 +190,16 @@ jobs:
       copy_to_acr: false
     secrets: inherit
 
+  # ============================================================================
+  # FRONTEND IMAGE BUILD
+  # ============================================================================
+  frontend-image:
+    name: Frontend Image
+    uses: ./.github/workflows/build-frontend-image.yaml
+    with:
+      skip_change_detection: true
+    secrets: inherit
+
   # ============================================================================
   # Operator
   # ============================================================================
@@ -354,7 +364,7 @@ jobs:
 
   deploy-cleanup:
     if: always()
-    needs: [deploy-operator, deploy-test-vllm, deploy-test-sglang, deploy-test-trtllm]
+    needs: [deploy-operator, deploy-test-vllm, deploy-test-sglang, deploy-test-trtllm, deploy-test-gaie]
     runs-on: prod-default-small-v2
     steps:
     - uses: actions/checkout@v4
@@ -366,9 +376,37 @@ jobs:
         vcluster_name: ${{ needs.deploy-operator.outputs.vcluster_name }}
         vcluster_namespace: ${{ needs.deploy-operator.outputs.namespace }}
 
+  # ============================================================================
+  # GAIE DEPLOY TEST
+  # ============================================================================
+  deploy-test-gaie:
+    name: GAIE Deploy Test
+    runs-on: prod-default-small-v2
+    needs: [deploy-operator, frontend-image, vllm-pipeline]
+    timeout-minutes: 30
+    permissions:
+      contents: read
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955  # v4.3.0
+      - name: Run GAIE Deploy Test
+        id: deploy-test
+        uses: ./.github/actions/dynamo-deploy-test
+        with:
+          kubeconfig_base64: ${{ secrets.AZURE_AKS_CI_KUBECONFIG_B64 }}
+          namespace: ${{ needs.deploy-operator.outputs.namespace }}
+          registry: ${{ secrets.AZURE_ACR_HOSTNAME }}
+          operator_tag: ${{ needs.deploy-operator.outputs.operator_tag }}
+          hf_token: ${{ secrets.HF_TOKEN }}
+          test_name: gaie
+          extra_pytest_args: >-
+            -m framework_with_gaie
+            --frontend-image=${{ secrets.AZURE_ACR_HOSTNAME }}/ai-dynamo/dynamo:${{ github.sha }}-frontend
+            --image=${{ secrets.AZURE_ACR_HOSTNAME }}/ai-dynamo/dynamo:${{ github.sha }}-vllm-runtime-cuda12-amd64
+
   deploy-status-check:
     runs-on: ubuntu-latest
-    needs: [deploy-operator, deploy-test-vllm, deploy-test-sglang, deploy-test-trtllm]
+    needs: [deploy-operator, deploy-test-vllm, deploy-test-sglang, deploy-test-trtllm, deploy-test-gaie]
     if: always()
     steps:
       - name: "Check all deploy test jobs"
@@ -383,7 +421,7 @@ jobs:
     name: Clean K8s builder if exists
     runs-on: prod-default-small-v2
     if: always()
-    needs: [vllm-pipeline, sglang-pipeline, trtllm-pipeline, vllm-dev-pipeline, sglang-dev-pipeline, trtllm-dev-pipeline, vllm-efa-pipeline, trtllm-efa-pipeline, operator]
+    needs: [vllm-pipeline, sglang-pipeline, trtllm-pipeline, vllm-dev-pipeline, sglang-dev-pipeline, trtllm-dev-pipeline, vllm-efa-pipeline, trtllm-efa-pipeline, operator, frontend-image]
     steps:
     - name: Checkout repository
       uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955  # v4.3.0
@@ -404,7 +442,7 @@ jobs:
     name: Notify Slack
     runs-on: prod-builder-amd-v1
     if: always() && failure()
-    needs: [ vllm-pipeline, sglang-pipeline, trtllm-pipeline, vllm-dev-pipeline, sglang-dev-pipeline, trtllm-dev-pipeline, vllm-efa-pipeline, trtllm-efa-pipeline, operator, deploy-operator, deploy-test-vllm, deploy-test-sglang, deploy-test-trtllm ]
+    needs: [ vllm-pipeline, sglang-pipeline, trtllm-pipeline, vllm-dev-pipeline, sglang-dev-pipeline, trtllm-dev-pipeline, vllm-efa-pipeline, trtllm-efa-pipeline, operator, frontend-image, deploy-operator, deploy-test-vllm, deploy-test-sglang, deploy-test-trtllm, deploy-test-gaie ]
     permissions:
       contents: read
     steps:

.github/workflows/shared-deploy-test-framework.yml

@@ -85,3 +85,4 @@ jobs:
           profile: ${{ matrix.profile }}
           image: ${{ secrets.AZURE_ACR_HOSTNAME }}/ai-dynamo/dynamo:${{ github.sha }}-${{ inputs.image_suffix }}
           platform_arch: amd64
+          extra_pytest_args: -m framework_only

docs/kubernetes/inference-gateway.md

@@ -122,7 +122,11 @@ For the HttpRoute service make sure to specify the namespace where your gateway
 ```bash
 cd <dynamo-source-root>
 # kubectl get httproutes -n my-model # Make sure you do not have an incompatible HttpRoute running, delete if so.
+# Choose disagg or agg example
+kubectl apply -f examples/backends/vllm/deploy/gaie/disagg.yaml -n my-model
+# or
 kubectl apply -f examples/backends/vllm/deploy/gaie/agg.yaml -n my-model
+# make sure to apply the route
 kubectl apply -f examples/backends/vllm/deploy/gaie/http-route.yaml -n my-model
 ```
 

examples/backends/vllm/deploy/gaie/agg.yaml

@@ -3,7 +3,7 @@
 apiVersion: nvidia.com/v1alpha1
 kind: DynamoGraphDeployment
 metadata:
-  name: qwen-agg
+  name: qwen
 spec:
   backendFramework: vllm
   services:

examples/backends/vllm/deploy/gaie/disagg.yaml

+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: nvidia.com/v1alpha1
+kind: DynamoGraphDeployment
+metadata:
+  name: qwen
+spec:
+  backendFramework: vllm
+  services:
+    Epp:
+      envFromSecret: hf-token-secret
+      componentType: epp
+      replicas: 1
+      extraPodSpec:
+        mainContainer:
+          image: nvcr.io/nvidia/ai-dynamo/epp-image:my-tag
+          imagePullPolicy: IfNotPresent
+          env:
+            - name: DYN_KV_CACHE_BLOCK_SIZE
+              value: "16"
+            - name: DYN_MODEL_NAME
+              value: "Qwen/Qwen3-0.6B"
+            - name: DYN_ENFORCE_DISAGG
+              value: "true"
+      eppConfig:
+        config:
+          plugins:
+            - type: disagg-profile-handler
+            - name: prefill-filter
+              type: label-filter
+              parameters:
+                label: "nvidia.com/dynamo-sub-component-type"
+                validValues:
+                  - "prefill"
+                allowsNoLabel: false
+            - name: decode-filter
+              type: label-filter
+              parameters:
+                label: "nvidia.com/dynamo-sub-component-type"
+                validValues:
+                  - "decode"
+                allowsNoLabel: false
+            - name: picker
+              type: max-score-picker
+            - name: dyn-prefill
+              type: dyn-prefill-scorer
+            - name: dyn-decode
+              type: dyn-decode-scorer
+          schedulingProfiles:
+          - name: prefill
+            plugins:
+            - pluginRef: prefill-filter
+              weight: 1
+            - pluginRef: dyn-prefill
+              weight: 1
+            - pluginRef: picker
+              weight: 1
+          - name: decode
+            plugins:
+            - pluginRef: decode-filter
+              weight: 1
+            - pluginRef: dyn-decode
+              weight: 1
+            - pluginRef: picker
+              weight: 1
+    VllmPrefillWorker:
+      componentType: worker
+      subComponentType: prefill
+      envFromSecret: hf-token-secret
+      sharedMemory:
+        size: 2Gi
+      frontendSidecar:
+        image: docker.io/lambda108/dynamo:post-rebase
+        args:
+          - -m
+          - dynamo.frontend
+          - --router-mode
+          - direct
+        envFromSecret: hf-token-secret
+      extraPodSpec:
+        affinity:
+          podAffinity:
+            preferredDuringSchedulingIgnoredDuringExecution:
+            - weight: 100
+              podAffinityTerm:
+                labelSelector:
+                  matchExpressions:
+                  - key: nvidia.com/dynamo-component-type
+                    operator: In
+                    values:
+                    - worker
+                topologyKey: kubernetes.io/hostname
+        mainContainer:
+          env:
+            - name: SERVED_MODEL_NAME
+              value: "Qwen/Qwen3-0.6B"
+            - name: MODEL_PATH
+              value: "Qwen/Qwen3-0.6B"
+          args:
+          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 1 --data-parallel-size 1 --disaggregation-mode prefill --kv-transfer-config '{\"kv_connector\":\"NixlConnector\",\"kv_role\":\"kv_both\"}' --gpu-memory-utilization 0.90 --enable-prefix-caching --block-size 16 --kv-events-config '{\"enable_kv_cache_events\":true}'"
+          command:
+          - /bin/sh
+          - -c
+          image: docker.io/lambda108/dynamo:post-rebase
+          imagePullPolicy: IfNotPresent
+          workingDir: /workspace/examples/backends/vllm
+      replicas: 1
+      resources:
+        limits:
+          gpu: "1"
+        requests:
+          gpu: "1"
+    VllmDecodeWorker:
+      componentType: worker
+      subComponentType: decode
+      envFromSecret: hf-token-secret
+      sharedMemory:
+        size: 2Gi
+      frontendSidecar:
+        image: docker.io/lambda108/dynamo:post-rebase
+        args:
+          - -m
+          - dynamo.frontend
+          - --router-mode
+          - direct
+        envFromSecret: hf-token-secret
+      extraPodSpec:
+        affinity:
+          podAffinity:
+            preferredDuringSchedulingIgnoredDuringExecution:
+            - weight: 100
+              podAffinityTerm:
+                labelSelector:
+                  matchExpressions:
+                  - key: nvidia.com/dynamo-component-type
+                    operator: In
+                    values:
+                    - worker
+                topologyKey: kubernetes.io/hostname
+        mainContainer:
+          env:
+            - name: SERVED_MODEL_NAME
+              value: "Qwen/Qwen3-0.6B"
+            - name: MODEL_PATH
+              value: "Qwen/Qwen3-0.6B"
+          args:
+          - "python3 -m dynamo.vllm --model $MODEL_PATH --served-model-name $SERVED_MODEL_NAME --tensor-parallel-size 1 --data-parallel-size 1 --kv-transfer-config '{\"kv_connector\":\"NixlConnector\",\"kv_role\":\"kv_both\"}' --gpu-memory-utilization 0.90 --enable-prefix-caching --block-size 16 --kv-events-config '{\"enable_kv_cache_events\":true}'"
+          command:
+          - /bin/sh
+          - -c
+          image: docker.io/lambda108/dynamo:post-rebase
+          imagePullPolicy: IfNotPresent
+          workingDir: /workspace/examples/backends/vllm
+      replicas: 1
+      resources:
+        limits:
+          gpu: "1"
+        requests:
+          gpu: "1"
\ No newline at end of file

examples/backends/vllm/deploy/gaie/http-route.yaml

@@ -18,18 +18,18 @@
 apiVersion: gateway.networking.k8s.io/v1
 kind: HTTPRoute
 metadata:
-  name: qwen-agg-route
+  name: qwen-route
 spec:
   parentRefs:
     - group: gateway.networking.k8s.io
       kind: Gateway
       name: inference-gateway
-      namespace: my-model # the namespace where your gateway is deployed.
+      namespace: default # the namespace where your gateway is deployed.
   rules:
     - backendRefs:
         - group: inference.networking.k8s.io
           kind: InferencePool
-          name: qwen-agg-pool
+          name: qwen-pool
           port: 8000
           weight: 1
       matches:

pyproject.toml

@@ -254,6 +254,8 @@ markers = [
     "k8s: marks tests as requiring Kubernetes",
     "fault_tolerance: marks tests as fault tolerance tests",
     "deploy: marks tests as deployment tests",
+    "framework_only: marks standard framework deployment tests (vllm, sglang, trtllm)",
+    "framework_with_gaie: marks tests for GAIE (Gateway API Inference Extension) deployment",
     # Built-in markers
     "skip: skip this test",
     "skipif: skip if condition is true",

recipes/llama-3-70b/vllm/disagg-single-node/gaie/deploy.yaml

@@ -22,6 +22,8 @@ spec:
               value: "128"
             - name: DYN_MODEL_NAME
               value: "RedHatAI/Llama-3.3-70B-Instruct-FP8-dynamic"
+            - name: DYN_ENFORCE_DISAGG
+              value: "true"
       eppConfig:
         config:
           plugins:
@@ -81,6 +83,10 @@ spec:
           - direct
         envFromSecret: hf-token-secret
       extraPodSpec:
+        tolerations:
+          - key: nvidia.com/gpu
+            operator: Exists
+            effect: NoSchedule
         affinity:
           podAffinity:
             preferredDuringSchedulingIgnoredDuringExecution:
@@ -132,6 +138,10 @@ spec:
           - direct
         envFromSecret: hf-token-secret
       extraPodSpec:
+        tolerations:
+          - key: nvidia.com/gpu
+            operator: Exists
+            effect: NoSchedule
         affinity:
           podAffinity:
             preferredDuringSchedulingIgnoredDuringExecution:

tests/deploy/conftest.py

@@ -40,6 +40,12 @@ def pytest_addoption(parser: pytest.Parser) -> None:
         help="Deployment profile to test (e.g., agg, disagg, disagg_router). "
         "If not specified, runs all profiles for the selected framework.",
     )
+    parser.addoption(
+        "--frontend-image",
+        type=str,
+        default=None,
+        help="Frontend container image (used by GAIE deploy tests).",
+    )
 
 
 @dataclass(frozen=True)

tests/deploy/test_deploy.py

@@ -9,14 +9,23 @@ to chat completion requests correctly.
 """
 
 import logging
+import os
+import subprocess
+import time
 from typing import Any, Dict
 
+import kr8s
 import pytest
 import requests
+import yaml
 
 from tests.deploy.conftest import DeploymentTarget
 from tests.utils.client import send_request, wait_for_model_availability
-from tests.utils.managed_deployment import DeploymentSpec, ManagedDeployment
+from tests.utils.managed_deployment import (
+    DeploymentSpec,
+    ManagedDeployment,
+    _get_workspace_dir,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -38,6 +47,7 @@ DEFAULT_REQUEST_TIMEOUT = 120
 # Minimum response content length to validate that the model is generating meaningful output.
 # This matches the validation threshold from the original shell-based deployment tests.
 MIN_RESPONSE_CONTENT_LENGTH = 100
+GAIE_MODEL_NAME = "Qwen/Qwen3-0.6B"
 
 
 def validate_chat_response(
@@ -100,6 +110,7 @@ def validate_chat_response(
     return data
 
 
+@pytest.mark.framework_only
 @pytest.mark.k8s
 @pytest.mark.deploy
 @pytest.mark.post_merge
@@ -213,3 +224,203 @@ async def test_deployment(
             f"Deployment test PASSED for {deployment_target.test_id} "
             f"(source: {deployment_target.source}, model: {model}, namespace: {namespace})"
         )
+
+
+# GAIE (Gateway API Inference Extension) deployment test
+@pytest.mark.framework_with_gaie
+@pytest.mark.k8s
+@pytest.mark.deploy
+@pytest.mark.post_merge
+@pytest.mark.e2e
+@pytest.mark.timeout(900)
+async def test_gaie_deployment(
+    image: str,
+    namespace: str,
+    skip_service_restart: bool,
+    request,
+) -> None:
+    """Test GAIE disaggregated deployment with vLLM workers.
+
+    Applies the GAIE DynamoGraphDeployment (with CI-built images) and the
+    companion HTTPRoute, then verifies inference works end-to-end through
+    the full Gateway path.
+    """
+    frontend_image = request.config.getoption("--frontend-image")
+    worker_image = image
+
+    assert frontend_image, "--frontend-image is required for GAIE deploy test"
+    assert worker_image, "--image is required for GAIE deploy test"
+    assert namespace, "--namespace is required for GAIE deploy test"
+
+    workspace = _get_workspace_dir()
+    gaie_dir = os.path.join(workspace, "examples", "backends", "vllm", "deploy", "gaie")
+    disagg_path = os.path.join(gaie_dir, "disagg.yaml")
+    httproute_path = os.path.join(gaie_dir, "http-route.yaml")
+
+    assert os.path.exists(disagg_path), f"disagg.yaml not found: {disagg_path}"
+    assert os.path.exists(
+        httproute_path
+    ), f"http-route.yaml not found: {httproute_path}"
+
+    deployment_spec = DeploymentSpec(disagg_path)
+    deployment_spec.namespace = namespace
+
+    logger.info(f"Frontend image: {frontend_image}")
+    logger.info(f"Worker image: {worker_image}")
+
+    deployment_spec.set_image(frontend_image, service_name="Epp")
+    for worker in ("VllmPrefillWorker", "VllmDecodeWorker"):
+        deployment_spec.set_image(worker_image, service_name=worker)
+        deployment_spec.set_frontend_sidecar_image(frontend_image, service_name=worker)
+
+    route_hostname = f"{namespace}.example.com"
+    logger.info(f"HTTPRoute hostname: {route_hostname}")
+
+    with open(httproute_path) as f:
+        httproute_spec = yaml.safe_load(f)
+    httproute_spec["spec"]["hostnames"] = [route_hostname]
+    httproute_yaml = yaml.safe_dump(httproute_spec)
+
+    logger.info("Applying GAIE HTTPRoute...")
+    result = subprocess.run(
+        ["kubectl", "apply", "-n", namespace, "-f", "-"],
+        input=httproute_yaml,
+        capture_output=True,
+        text=True,
+    )
+    logger.info(f"HTTPRoute apply stdout: {result.stdout}")
+    if result.stderr:
+        logger.warning(f"HTTPRoute apply stderr: {result.stderr}")
+    assert result.returncode == 0, f"Failed to apply HTTPRoute: {result.stderr}"
+
+    # Debug: verify namespace state before creating DGD
+    logger.info(f"Namespace: {namespace}")
+    ns_check = subprocess.run(
+        ["kubectl", "get", "namespace", namespace],
+        capture_output=True,
+        text=True,
+    )
+    logger.info(f"Namespace check: {ns_check.stdout.strip()}")
+    if ns_check.returncode != 0:
+        logger.error(f"Namespace not found: {ns_check.stderr}")
+
+    # Debug: check if operator CRD is registered
+    crd_check = subprocess.run(
+        ["kubectl", "get", "crd", "dynamographdeployments.nvidia.com"],
+        capture_output=True,
+        text=True,
+    )
+    logger.info(f"CRD check: {crd_check.stdout.strip()}")
+    if crd_check.returncode != 0:
+        logger.error(f"CRD not found: {crd_check.stderr}")
+
+    # Debug: check operator pod status
+    operator_check = subprocess.run(
+        [
+            "kubectl",
+            "get",
+            "pods",
+            "-n",
+            namespace,
+            "-l",
+            "app.kubernetes.io/name=dynamo-operator",
+        ],
+        capture_output=True,
+        text=True,
+    )
+    logger.info(f"Operator pods: {operator_check.stdout.strip()}")
+
+    # Debug: log the full deployment spec being submitted
+    logger.info(f"DGD name: {deployment_spec.name}")
+    logger.info(f"DGD namespace: {deployment_spec.namespace}")
+    logger.info(f"DGD services: {[s.name for s in deployment_spec.services]}")
+
+    async with ManagedDeployment(
+        log_dir=request.node.name,
+        deployment_spec=deployment_spec,
+        namespace=namespace,
+        skip_service_restart=skip_service_restart,
+        frontend_service_name="Epp",
+    ) as deployment:
+        # Debug: check what DGDs exist after creation
+        dgd_check = subprocess.run(
+            ["kubectl", "get", "dynamographdeployments", "-n", namespace],
+            capture_output=True,
+            text=True,
+        )
+        logger.info(f"DGDs after creation: {dgd_check.stdout.strip()}")
+
+        pod_check = subprocess.run(
+            ["kubectl", "get", "pods", "-n", namespace, "-o", "wide"],
+            capture_output=True,
+            text=True,
+        )
+        logger.info(f"Pods after creation: {pod_check.stdout.strip()}")
+        epp_pods = deployment.get_pods(["Epp"])
+        epp_pod_list = epp_pods.get("Epp", [])
+        assert len(epp_pod_list) > 0, "No EPP pods found for GAIE deployment"
+        logger.info(f"Found EPP pod: {epp_pod_list[0].name}")
+
+        gateway_svcs = list(
+            kr8s.get("services", "inference-gateway", namespace=namespace)
+        )
+        assert (
+            len(gateway_svcs) > 0
+        ), f"inference-gateway service not found in namespace {namespace}"
+        gateway_pf = gateway_svcs[0].portforward(remote_port=80, local_port=0)
+        gateway_pf.start()
+        time.sleep(2)
+
+        try:
+            gateway_url = f"http://localhost:{gateway_pf.local_port}"
+            logger.info(f"Gateway port-forward established: {gateway_url}")
+
+            endpoint = deployment_spec.endpoint
+            headers = {"Host": route_hostname}
+            logger.info(f"Using Host header: {route_hostname}")
+
+            model_ready = wait_for_model_availability(
+                url=gateway_url,
+                endpoint=endpoint,
+                model=GAIE_MODEL_NAME,
+                logger=logger,
+                max_attempts=30,
+                headers=headers,
+            )
+            assert model_ready, (
+                f"Model '{GAIE_MODEL_NAME}' did not become available "
+                f"within the timeout period"
+            )
+
+            url = f"{gateway_url}{endpoint}"
+            payload = {
+                "model": GAIE_MODEL_NAME,
+                "messages": [{"role": "user", "content": TEST_PROMPT}],
+                "max_tokens": DEFAULT_MAX_TOKENS,
+                "temperature": DEFAULT_TEMPERATURE,
+                "stream": False,
+            }
+            logger.info(f"Sending inference request to {url}")
+            response = requests.post(
+                url,
+                json=payload,
+                headers=headers,
+                timeout=DEFAULT_REQUEST_TIMEOUT,
+            )
+
+            validate_chat_response(
+                response=response,
+                expected_model=GAIE_MODEL_NAME,
+                min_content_length=MIN_RESPONSE_CONTENT_LENGTH,
+            )
+
+            data = response.json()
+            content = data["choices"][0]["message"]["content"]
+            logger.info(
+                f"GAIE deployment test PASSED | "
+                f"model={data['model']}, status={response.status_code}, "
+                f"response_length={len(content)} chars\n"
+                f"Model response: {content}"
+            )
+        finally:
+            gateway_pf.stop()

tests/utils/client.py

@@ -159,6 +159,7 @@ def wait_for_model_availability(
     logger: logging.Logger,
     max_attempts: int = 15,
     attempt_timeouts: list[float] | None = None,
+    headers: dict[str, str] | None = None,
 ) -> bool:
     """
     Wait for model to be available by sending test requests.
@@ -197,7 +198,9 @@ def wait_for_model_availability(
             logger.debug(
                 f"Testing model availability at {test_url} (attempt {attempt+1}/{max_attempts}, timeout={timeout_val}s)"
             )
-            response = requests.post(test_url, json=test_payload, timeout=timeout_val)
+            response = requests.post(
+                test_url, json=test_payload, timeout=timeout_val, headers=headers
+            )
 
             if response.status_code == 200:
                 logger.info(f"Model '{model}' is available and responding")

tests/utils/managed_deployment.py

@@ -66,6 +66,20 @@ class ServiceSpec:
             self._spec["extraPodSpec"]["mainContainer"] = {}
         self._spec["extraPodSpec"]["mainContainer"]["image"] = value
 
+    @property
+    def frontend_sidecar_image(self) -> Optional[str]:
+        """Container image for the frontendSidecar (if present)."""
+        try:
+            return self._spec["frontendSidecar"]["image"]
+        except KeyError:
+            return None
+
+    @frontend_sidecar_image.setter
+    def frontend_sidecar_image(self, value: str):
+        if "frontendSidecar" not in self._spec:
+            self._spec["frontendSidecar"] = {}
+        self._spec["frontendSidecar"]["image"] = value
+
     @property
     def envs(self) -> list[dict[str, str]]:
         """Environment variables for the service"""
@@ -230,6 +244,16 @@ class DeploymentSpec:
         for service in services:
             service.image = image
 
+    def set_frontend_sidecar_image(
+        self, image: str, service_name: Optional[str] = None
+    ):
+        if service_name is None:
+            services = self.services
+        else:
+            services = [self[service_name]]
+        for service in services:
+            service.frontend_sidecar_image = image
+
     def set_tensor_parallel(self, tp_size: int, service_names: Optional[list] = None):
         """Scale deployment for different tensor parallel configurations