tanmayv25 · Sep 18, 2026 · Sep 9, 2026 · Sep 9, 2026 · Sep 9, 2026 · Sep 9, 2026
@@ -344,6 +344,7 @@ sidecar:
   # acceptable CI fan-out is agreed. Keep the remaining scope narrow for now.
   - 'lib/sidecar/**'
   - '.github/workflows/shared-build-sidecar.yml'
+  - 'tests/serve/test_sidecar.py'
   # Compliance policy and baseline changes can fail the sidecar image build.
   - 'container/compliance/**'
   - '!**/*.md'

@@ -815,7 +815,13 @@ jobs:
       enable_coverage: true
       run_cpu_only_tests: true
       cpu_only_test_markers: vllm and gpu_0
-      gpu_test_markers: vllm and gpu_1
+      # not sidecar: sidecar-marked tests need the dynamo-vllm-sidecar binary
+      # installed (see pr.yaml's sidecar-vllm-test), which this plain
+      # vllm-runtime image doesn't have.
+      # TODO: this ad-hoc exclusion is duplicated across pr.yaml, post-merge-ci.yml,
+      # and nightly-ci.yml's gpu_test_markers; consolidate into shared-test.yml
+      # once there's a second marker that needs the same treatment.
+      gpu_test_markers: vllm and gpu_1 and not sidecar
       gpu_test_timeout_minutes: 240
       # Profiled tests run in the parallel stage; unprofiled fall through to sequential.
       # 24 GiB admits all currently profiled vLLM tests (max is ~20.4 GiB) on a 48 GiB GPU.
@@ -887,7 +893,7 @@ jobs:
       platform: '["amd64"]'
       enable_coverage: true
       run_sanity_check: false
-      gpu_test_markers: vllm and (gpu_1 or gpu_2)
+      gpu_test_markers: vllm and (gpu_1 or gpu_2) and not sidecar
       gpu_test_timeout_minutes: 240
       # 80 GiB H100 admits more of the profiled pool per slot than the 24 GiB lanes.
       run_gpu_parallel_tests: true
@@ -914,7 +920,7 @@ jobs:
       run_cpu_only_tests: true
       cpu_only_test_markers: sglang and gpu_0
       # Exclude h100 tests: they need >24 GiB and run on the sglang-h100-test lane below.
-      gpu_test_markers: sglang and gpu_1 and not h100
+      gpu_test_markers: sglang and gpu_1 and not h100 and not sidecar
       gpu_test_timeout_minutes: 240
       # Profiled tests run in the VRAM-aware GPU stage; unprofiled fall through to sequential.
       # Current single-GPU runners are 24 GiB, so this cap admits the profiled SGLang pool
@@ -986,7 +992,7 @@ jobs:
       platform: '["amd64"]'
       enable_coverage: true
       run_sanity_check: false
-      gpu_test_markers: sglang and (gpu_1 or gpu_2)
+      gpu_test_markers: sglang and (gpu_1 or gpu_2) and not sidecar
       gpu_test_timeout_minutes: 240
       # 80 GiB H100 admits more of the profiled pool per slot than the 24 GiB lanes.
       run_gpu_parallel_tests: true
@@ -1012,7 +1018,7 @@ jobs:
       enable_coverage: true
       run_cpu_only_tests: true
       cpu_only_test_markers: trtllm and gpu_0
-      gpu_test_markers: trtllm and gpu_1
+      gpu_test_markers: trtllm and gpu_1 and not sidecar
       gpu_test_timeout_minutes: 240
       # Profiled tests run in the VRAM-aware GPU stage; unprofiled fall through to sequential.
       # Current single-GPU runners are 24 GiB, so this cap admits the profiled TRT-LLM pool
@@ -1062,7 +1068,7 @@ jobs:
       platform: '["amd64"]'
       enable_coverage: true
       run_sanity_check: false
-      gpu_test_markers: trtllm and (gpu_1 or gpu_2)
+      gpu_test_markers: trtllm and (gpu_1 or gpu_2) and not sidecar
       gpu_test_timeout_minutes: 240
       # 80 GiB H100 admits more of the profiled pool per slot than the 24 GiB lanes.
       run_gpu_parallel_tests: true

@@ -214,6 +214,10 @@ jobs:
       extra_tags: |
         ${{ github.ref_name == 'main' && 'main-vllm-runtime' || '' }}
         ${{ github.ref_name == 'main' && format('main-vllm-runtime-{0}', github.sha) || '' }}
+      # Lets sidecar-vllm-test-image (pr.yaml) layer onto the latest main -test
+      # image instead of rebuilding vllm-runtime on every lib/sidecar/**-only PR.
+      test_extra_tags: |
+        ${{ github.ref_name == 'main' && 'main-vllm-runtime-test' || '' }}
       archive_sources: false
     secrets: inherit
 
@@ -279,6 +283,10 @@ jobs:
       extra_tags: |
         ${{ github.ref_name == 'main' && 'main-sglang-runtime' || '' }}
         ${{ github.ref_name == 'main' && format('main-sglang-runtime-{0}', github.sha) || '' }}
+      # Lets sidecar-sglang-test-image (pr.yaml) layer onto the latest main -test
+      # image instead of rebuilding sglang-runtime on every lib/sidecar/**-only PR.
+      test_extra_tags: |
+        ${{ github.ref_name == 'main' && 'main-sglang-runtime-test' || '' }}
       archive_sources: false
     secrets: inherit
 
@@ -344,6 +352,10 @@ jobs:
       extra_tags: |
         ${{ github.ref_name == 'main' && 'main-trtllm-runtime' || '' }}
         ${{ github.ref_name == 'main' && format('main-trtllm-runtime-{0}', github.sha) || '' }}
+      # Lets sidecar-trtllm-test-image (pr.yaml) layer onto the latest main -test
+      # image instead of rebuilding trtllm-runtime on every lib/sidecar/**-only PR.
+      test_extra_tags: |
+        ${{ github.ref_name == 'main' && 'main-trtllm-runtime-test' || '' }}
       archive_sources: false
     secrets: inherit
 
@@ -452,7 +464,10 @@ jobs:
       platform: '["amd64", "arm64"]' # arm64 for CPU tests, single GPU tests are skipped
       run_cpu_only_tests: true
       cpu_only_test_markers: '(pre_merge or post_merge) and vllm and gpu_0'
-      gpu_test_markers: '(pre_merge or post_merge) and vllm and gpu_1'
+      # not sidecar: sidecar-marked tests need dynamo-vllm-sidecar layered onto
+      # the image (see pr.yaml's sidecar-vllm-test), which this plain
+      # vllm-runtime image doesn't have.
+      gpu_test_markers: '(pre_merge or post_merge) and vllm and gpu_1 and not sidecar'
       gpu_test_timeout_minutes: 120
       # Profiled tests (those carrying @pytest.mark.profiled_vram_gib(N)) run concurrently in
       # the parallel stage; the rest fall through to the sequential stage. 24 GiB admits all
@@ -550,7 +565,10 @@ jobs:
       platform: '["amd64", "arm64"]' # arm64 for CPU tests, single GPU tests are skipped
       run_cpu_only_tests: true
       cpu_only_test_markers: '(pre_merge or post_merge) and sglang and gpu_0'
-      gpu_test_markers: '(pre_merge or post_merge) and sglang and gpu_1'
+      # not sidecar: sidecar-marked tests need dynamo-sglang-sidecar layered
+      # onto the image (see pr.yaml's sidecar-sglang-test), which this plain
+      # sglang-runtime image doesn't have.
+      gpu_test_markers: '(pre_merge or post_merge) and sglang and gpu_1 and not sidecar'
       gpu_test_timeout_minutes: 120
       # Profiled tests run in the VRAM-aware GPU stage; unprofiled fall through to sequential.
       # Current single-GPU runners are 24 GiB, so this cap admits the profiled SGLang pool
@@ -590,7 +608,10 @@ jobs:
       platform: '["amd64", "arm64"]' # arm64 for CPU tests, single GPU tests are skipped
       run_cpu_only_tests: true
       cpu_only_test_markers: '(pre_merge or post_merge) and trtllm and gpu_0'
-      gpu_test_markers: '(pre_merge or post_merge) and trtllm and gpu_1'
+      # not sidecar: sidecar-marked tests need dynamo-trtllm-sidecar layered
+      # onto the image (see pr.yaml's sidecar-trtllm-test), which this plain
+      # trtllm-runtime image doesn't have.
+      gpu_test_markers: '(pre_merge or post_merge) and trtllm and gpu_1 and not sidecar'
       gpu_test_timeout_minutes: 120
       # Profiled tests run in the VRAM-aware GPU stage; unprofiled fall through to sequential.
       # Current single-GPU runners are 24 GiB, so this cap admits the profiled TRT-LLM pool

@@ -133,6 +133,12 @@
       - snapshot-placeholder-trtllm
       - power-agent
       - sidecar-build
+      - sidecar-vllm-test-image
+      - sidecar-sglang-test-image
+      - sidecar-trtllm-test-image
+      - sidecar-vllm-test
+      - sidecar-sglang-test
+      - sidecar-trtllm-test
       - vllm-build
       - vllm-local-dev-build
       - sglang-local-dev-build
@@ -358,14 +364,87 @@
       diff_base_branch: ${{ needs.changed-files.outputs.base_ref }}
       diff_merge_base_sha: ${{ needs.changed-files.outputs.merge_base_sha }}
 
+  # --------------------------------------------------------------------------
+  # Sidecar E2E: lib/sidecar/*/launch/agg.sh, run against the backend's own
+  # unmodified -test image (no CI-only derived image/extra push). shared-test.yml
+  # downloads the dynamo-*-sidecar binary sidecar-build already published as a
+  # workflow artifact and installs it via PATH before pytest runs. When the
+  # backend itself wasn't rebuilt (a lib/sidecar/**-only PR), source_ref falls
+  # back to the floating `main` tag for that backend's -test image. sidecar-build
+  # is gated identically to these jobs, so `needs.sidecar-build.result` is a
+  # genuine success/failure signal here, not a masked skip. gpu_1 fan-out is a
+  # deliberate exception to the narrow-scope TODO on the `sidecar` filter in
+  # .github/filters.yaml; revisit if CI cost becomes an issue.
+  # --------------------------------------------------------------------------
+
+  sidecar-vllm-test:
+    name: sidecar-runtime # This name overlaps with other sidecar jobs to group them in the UI
+    needs: [changed-files, vllm-build, sidecar-build]
+    if: ${{ !cancelled() && needs.changed-files.outputs.sidecar == 'true' && needs.sidecar-build.result == 'success' }}
+    uses: $/.github/workflows/shared-test.yml
+    with:
+      test_suite_name: vllm
+      test_type: Sidecar E2E Test
+      amd_runner: prod-tester-amd-gpu-v2
+      target_tag_plain: ${{ needs.vllm-build.result == 'success' && needs.vllm-build.outputs.target_tag_plain || 'vllm-runtime' }}
+      source_ref: ${{ needs.vllm-build.result == 'success' && github.sha || 'main' }}
+      cuda_version: '["13.0"]'
+      platform: '["amd64"]'
+      run_sanity_check: false
+      gpu_test_markers: pre_merge and sidecar and vllm and gpu_1
+      gpu_test_timeout_minutes: 45
+      sidecar_backend: vllm
+      sidecar_binary_artifact: ${{ needs.sidecar-build.outputs.binary_artifact }}
+    secrets: inherit
+
+  sidecar-sglang-test:
+    name: sidecar-runtime # This name overlaps with other sidecar jobs to group them in the UI
+    needs: [changed-files, sglang-build, sidecar-build]
+    if: ${{ !cancelled() && needs.changed-files.outputs.sidecar == 'true' && needs.sidecar-build.result == 'success' }}
+    uses: $/.github/workflows/shared-test.yml
+    with:
+      test_suite_name: sglang
+      test_type: Sidecar E2E Test
+      amd_runner: prod-tester-amd-gpu-v2
+      target_tag_plain: ${{ needs.sglang-build.result == 'success' && needs.sglang-build.outputs.target_tag_plain || 'sglang-runtime' }}
+      source_ref: ${{ needs.sglang-build.result == 'success' && github.sha || 'main' }}
+      cuda_version: '["13.0"]'
+      platform: '["amd64"]'
+      run_sanity_check: false
+      gpu_test_markers: pre_merge and sidecar and sglang and gpu_1
+      gpu_test_timeout_minutes: 45
+      sidecar_backend: sglang
+      sidecar_binary_artifact: ${{ needs.sidecar-build.outputs.binary_artifact }}
+    secrets: inherit
+
+  sidecar-trtllm-test:
+    name: sidecar-runtime # This name overlaps with other sidecar jobs to group them in the UI
+    needs: [changed-files, trtllm-build, sidecar-build]
+    if: ${{ !cancelled() && needs.changed-files.outputs.sidecar == 'true' && needs.sidecar-build.result == 'success' }}
+    uses: $/.github/workflows/shared-test.yml
+    with:
+      test_suite_name: trtllm
+      test_type: Sidecar E2E Test
+      amd_runner: prod-tester-amd-gpu-v2
+      target_tag_plain: ${{ needs.trtllm-build.result == 'success' && needs.trtllm-build.outputs.target_tag_plain || 'trtllm-runtime' }}
+      source_ref: ${{ needs.trtllm-build.result == 'success' && github.sha || 'main' }}
+      cuda_version: '["13.1"]'
+      platform: '["amd64"]'
+      run_sanity_check: false
+      gpu_test_markers: pre_merge and sidecar and trtllm and gpu_1
+      gpu_test_timeout_minutes: 45
+      sidecar_backend: trtllm
+      sidecar_binary_artifact: ${{ needs.sidecar-build.outputs.binary_artifact }}
+    secrets: inherit
+
 # ============================================================================
 # FRAMEWORK PIPELINES (Build → Test → Copy)
 # ============================================================================
  # ============================================================================
  # BUILD PIPELINES
  # ============================================================================

  vllm-build:
    name: vllm-runtime # This name overlaps with other vllm jobs to group them in the UI
    needs: [changed-files]
    # `sample` is here because sample tests piggyback on this image and run as part of vllm-test.
@@ -795,7 +874,10 @@
       # Keep the outer step budget above the test's 450s timeout so pytest can
       # finish teardown and upload diagnostics instead of being killed first.
       cpu_e2e_test_timeout_minutes: 15
-      gpu_test_markers: pre_merge and vllm and gpu_1
+      # not sidecar: sidecar-marked tests need dynamo-vllm-sidecar layered onto
+      # the image (see sidecar-vllm-test below), which this plain vllm-runtime
+      # image doesn't have.
+      gpu_test_markers: pre_merge and vllm and gpu_1 and not sidecar
       gpu_test_timeout_minutes: 45
       # Profiled tests run in the parallel stage; unprofiled fall through to sequential.
       # 24 GiB admits all currently profiled vLLM tests (max is ~20.4 GiB) on a 48 GiB GPU.
@@ -838,7 +920,10 @@
       platform: '["amd64", "arm64"]' # arm64 for CPU tests, single GPU tests are skipped
       run_cpu_only_tests: true
       cpu_only_test_markers: pre_merge and sglang and gpu_0
-      gpu_test_markers: pre_merge and sglang and gpu_1
+      # not sidecar: sidecar-marked tests need dynamo-sglang-sidecar layered
+      # onto the image (see sidecar-sglang-test below), which this plain
+      # sglang-runtime image doesn't have.
+      gpu_test_markers: pre_merge and sglang and gpu_1 and not sidecar
       # Profiled tests run in the VRAM-aware GPU stage; unprofiled fall through to sequential.
       # Current single-GPU runners are 24 GiB, so this cap admits the profiled SGLang pool
       # while yielding one auto slot today. Larger runners will get more slots from the same markers.
@@ -881,7 +966,10 @@
       platform: '["amd64", "arm64"]' # arm64 for CPU tests, single GPU tests are skipped
       run_cpu_only_tests: true
       cpu_only_test_markers: pre_merge and trtllm and gpu_0
-      gpu_test_markers: pre_merge and trtllm and gpu_1
+      # not sidecar: sidecar-marked tests need dynamo-trtllm-sidecar layered
+      # onto the image (see sidecar-trtllm-test below), which this plain
+      # trtllm-runtime image doesn't have.
+      gpu_test_markers: pre_merge and trtllm and gpu_1 and not sidecar
       # Profiled tests run in the VRAM-aware GPU stage; unprofiled fall through to sequential.
       # Current single-GPU runners are 24 GiB, so this cap admits the profiled TRT-LLM pool
       # while yielding one auto slot today. Larger runners will get more slots from the same markers.

@@ -37,6 +37,11 @@ on:
         required: false
         type: string
         default: ''
+      test_extra_tags:
+        description: 'Additional tags for the -test image only (newline-separated plain tag names, no ECR prefix). Unlike extra_tags (runtime image), the -test image otherwise gets no floating tag.'
+        required: false
+        type: string
+        default: ''
       no_cache:
         description: 'Disable Docker build cache'
         required: false
@@ -603,6 +608,14 @@ jobs:
       - name: Build and Push Test Image
         if: ${{ !inputs.compliance_only && inputs.target != 'dev' && inputs.target != 'local-dev' && inputs.target != 'frontend' }}
         shell: bash
+        env:
+          ECR_REGISTRY: ${{ vars.ECR_REGISTRY }}
+          ECR_REPOSITORY: ${{ vars.ECR_REPOSITORY }}
+          TEST_EXTRA_TAGS: ${{ inputs.test_extra_tags }}
+          # Routed through env (not inlined via ${{ }}) so a value containing
+          # shell metacharacters can't be interpreted as script rather than data.
+          IMAGE_URI: ${{ steps.calculate-target-tag.outputs.image_uri }}
+          TEST_IMAGE_URI: ${{ steps.calculate-target-tag.outputs.test_image_uri }}
         run: |
           PUSH_ARGS=""
           if [ "${{ inputs.push_image }}" == "true" ]; then
@@ -630,6 +643,15 @@ jobs:
             PYPI_SECRET_ARGS+=(--secret "id=pypi-netrc,src=${NETRC}")
           fi
 
+          TEST_TAG_ARGS=(-t "${TEST_IMAGE_URI}")
+          if [ -n "$TEST_EXTRA_TAGS" ]; then
+            while IFS= read -r tag; do
+              if [ -n "$tag" ]; then
+                TEST_TAG_ARGS+=(-t "${ECR_REGISTRY}/${ECR_REPOSITORY}:${tag}")
+              fi
+            done <<< "$TEST_EXTRA_TAGS"
+          fi
+
           docker buildx build \
             --progress=plain \
             ${PUSH_ARGS} \
@@ -638,8 +660,8 @@ jobs:
             --platform ${{ inputs.platform }} \
             --target ${TEST_TARGET} \
             -f container/Dockerfile.test \
-            --build-arg BASE_IMAGE=${{ steps.calculate-target-tag.outputs.image_uri }} \
-            -t ${{ steps.calculate-target-tag.outputs.test_image_uri }} .
+            --build-arg BASE_IMAGE="${IMAGE_URI}" \
+            "${TEST_TAG_ARGS[@]}" .
       - name: Show summary
         shell: bash
         if: ${{ inputs.push_image == 'true' && inputs.show_summary == 'true' }}

@@ -45,6 +45,13 @@ on:
         required: false
         type: string
         default: ''
+    outputs:
+      target_tag_plain:
+        description: 'Plain sidecar image tag prefix'
+        value: ${{ jobs.build.outputs.target_tag_plain }}
+      binary_artifact:
+        description: 'Name of the workflow artifact containing the amd64 dynamo-*-sidecar binaries'
+        value: ${{ jobs.build.outputs.binary_artifact }}
 
 jobs:
   build:
@@ -54,6 +61,9 @@ jobs:
     permissions:
       actions: read
       contents: read
+    outputs:
+      target_tag_plain: ${{ steps.calculate-target-tag.outputs.target_tag_plain }}
+      binary_artifact: ${{ steps.binary-artifact.outputs.name }}
     steps:
       - name: Checkout repository
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
@@ -123,6 +133,35 @@ jobs:
           no_load: true
           use_sccache: false
 
+      # No derived per-backend test image: shared-test.yml pulls the backend's
+      # own -test image unmodified and installs the sidecar binary from this
+      # artifact via PATH (see shared-test.yml's sidecar_binary_artifact
+      # input). amd64-only: sidecar tests only run on amd64 GPU runners.
+      - name: Extract sidecar binaries (amd64)
+        id: binary-artifact
+        shell: bash
+        env:
+          IMAGE_URI: ${{ steps.calculate-target-tag.outputs.image_uri }}
+        run: |
+          set -euo pipefail
+          docker pull --platform linux/amd64 "${IMAGE_URI}"
+          cid=$(docker create --platform linux/amd64 "${IMAGE_URI}")
+          trap 'docker rm -f "${cid}" >/dev/null 2>&1 || true' EXIT
+          mkdir -p sidecar-binaries
+          for backend in vllm sglang trtllm; do
+            docker cp "${cid}:/usr/local/bin/dynamo-${backend}-sidecar" \
+              "sidecar-binaries/dynamo-${backend}-sidecar"
+          done
+          name="dynamo-sidecar-binaries-${GITHUB_SHA}"
+          echo "name=${name}" >> "${GITHUB_OUTPUT}"
+
+      - name: Upload sidecar binaries
+        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f  # v6.0.0
+        with:
+          name: ${{ steps.binary-artifact.outputs.name }}
+          path: sidecar-binaries/
+          retention-days: 7
+
       - name: Resolve diff baseline
         id: resolve-diff-base
         if: ${{ inputs.diff_event_context != '' }}