Merge branch 'main' into jsm/autoscale

Author: julienmancuso

.github/actions/docker-build/action.yml

@@ -280,13 +280,6 @@ runs:
         if [ ${PARSER_EXIT_CODE} -eq 0 ] && [ -f "$COMPREHENSIVE_JSON" ]; then
           echo "✅ Comprehensive build metrics generated successfully"
           echo "📄 Output file: ${COMPREHENSIVE_JSON}"
-          echo ""
-          echo "=========================================="
-          echo "📋 FULL JSON OUTPUT (for debugging)"
-          echo "=========================================="
-          cat "$COMPREHENSIVE_JSON"
-          echo ""
-          echo "=========================================="
         else
           echo "⚠️  Metrics generation had issues but continuing..."
         fi

.github/actions/pytest/action.yml

@@ -54,7 +54,16 @@ runs:
         # Run pytest with detailed output and JUnit XML
         set +e  # Don't exit on test failures
 
-        docker run --runtime=nvidia --gpus all -w /workspace \
+        # Detect GPU availability and conditionally add GPU flags
+        GPU_FLAGS=""
+        if command -v nvidia-smi &> /dev/null && nvidia-smi &> /dev/null; then
+          echo "GPU detected, enabling GPU runtime"
+          GPU_FLAGS="--runtime=nvidia --gpus all"
+        else
+          echo "No GPU detected, running in CPU-only mode"
+        fi
+
+        docker run ${GPU_FLAGS} --rm -w /workspace \
           --cpus=${NUM_CPUS} \
           --network host \
           --name ${{ env.CONTAINER_ID }}_pytest \
@@ -130,4 +139,4 @@ runs:
         path: |
           test-results/pytest_test_report_${{ inputs.framework }}_${{ env.STR_TEST_TYPE }}_${{ inputs.platform_arch }}.xml
           test-results/test_metadata_${{ inputs.framework }}_${{ env.STR_TEST_TYPE }}_${{ inputs.platform_arch }}.json
-        retention-days: 7
+        retention-days: 7

.github/workflows/container-validation-backends.yml

@@ -179,23 +179,14 @@ jobs:
           azure_acr_user: ${{ secrets.AZURE_ACR_USER }}
           azure_acr_password: ${{ secrets.AZURE_ACR_PASSWORD }}
 
-      - name: Run unit tests
+      - name: Run tests
         if: ${{ matrix.platform.arch != 'arm64' }}
         uses: ./.github/actions/pytest
         with:
           image_tag: ${{ steps.build-image.outputs.image_tag }}
-          pytest_marks: "unit and vllm and gpu_1"
+          pytest_marks: "pre_merge and vllm"
           framework: "vllm"
-          test_type: "unit"
-          platform_arch: ${{ matrix.platform.arch }}
-      - name: Run e2e tests
-        if: ${{ matrix.platform.arch != 'arm64' }}
-        uses: ./.github/actions/pytest
-        with:
-          image_tag: ${{ steps.build-image.outputs.image_tag }}
-          pytest_marks: "e2e and vllm and gpu_1 and not slow"
-          framework: "vllm"
-          test_type: "e2e, gpu_1"
+          test_type: "pre_merge"
           platform_arch: ${{ matrix.platform.arch }}
 
   sglang:
@@ -246,23 +237,14 @@ jobs:
           azure_acr_user: ${{ secrets.AZURE_ACR_USER }}
           azure_acr_password: ${{ secrets.AZURE_ACR_PASSWORD }}
 
-      - name: Run unit tests
-        if: ${{ matrix.platform.arch != 'arm64' }}
-        uses: ./.github/actions/pytest
-        with:
-          image_tag: ${{ steps.build-image.outputs.image_tag }}
-          pytest_marks: "unit and sglang and gpu_1"
-          framework: "sglang"
-          test_type: "unit"
-          platform_arch: ${{ matrix.platform.arch }}
-      - name: Run e2e tests
+      - name: Run tests
         if: ${{ matrix.platform.arch != 'arm64' }}
         uses: ./.github/actions/pytest
         with:
           image_tag: ${{ steps.build-image.outputs.image_tag }}
-          pytest_marks: "e2e and sglang and gpu_1"
+          pytest_marks: "pre_merge and sglang"
           framework: "sglang"
-          test_type: "e2e, gpu_1"
+          test_type: "pre_merge"
           platform_arch: ${{ matrix.platform.arch }}
 
   trtllm:
@@ -313,23 +295,14 @@ jobs:
           azure_acr_user: ${{ secrets.AZURE_ACR_USER }}
           azure_acr_password: ${{ secrets.AZURE_ACR_PASSWORD }}
 
-      - name: Run unit tests
-        if: ${{ matrix.platform.arch != 'arm64' }}
-        uses: ./.github/actions/pytest
-        with:
-          image_tag: ${{ steps.build-image.outputs.image_tag }}
-          pytest_marks: "unit and trtllm_marker and gpu_1"
-          framework: "trtllm"
-          test_type: "unit"
-          platform_arch: ${{ matrix.platform.arch }}
-      - name: Run e2e tests
+      - name: Run tests
         if: ${{ matrix.platform.arch != 'arm64' }}
         uses: ./.github/actions/pytest
         with:
           image_tag: ${{ steps.build-image.outputs.image_tag }}
-          pytest_marks: "e2e and trtllm_marker and gpu_1 and not slow"
+          pytest_marks: "pre_merge and trtllm"
           framework: "trtllm"
-          test_type: "e2e, gpu_1"
+          test_type: "pre_merge"
           platform_arch: ${{ matrix.platform.arch }}
 
   deploy-test-fault-tolerance:

.github/workflows/container-validation-dynamo.yml

@@ -65,7 +65,7 @@ jobs:
           docker compose down
       - name: Run pytest (parallel tests with xdist)
         env:
-          PYTEST_MARKS: "pre_merge and parallel"
+          PYTEST_MARKS: "pre_merge and parallel and not (vllm or sglang or trtllm)"
         run: |
           docker run -w /workspace \
             --name ${{ env.CONTAINER_ID }}_pytest_parallel \
@@ -77,7 +77,7 @@ jobs:
           docker cp ${{ env.CONTAINER_ID }}_pytest_parallel:/workspace/${{ env.PYTEST_PARALLEL_XML_FILE }} . || echo "No parallel test report found"
       - name: Run pytest (sequential tests)
         env:
-          PYTEST_MARKS: "(pre_merge and not parallel) or mypy"
+          PYTEST_MARKS: "((pre_merge and not parallel) or mypy) and not (vllm or sglang or trtllm)"
         run: |
           docker run -w /workspace \
             --name ${{ env.CONTAINER_ID }}_pytest \