ci: adding timeouts (#6062)

.github/actions/skopeo-copy/action.yml

@@ -105,7 +105,20 @@ runs:
         echo "  Source: ${SOURCE_REF}"
         echo "  Target: ${TARGET_REF}"
 
-        skopeo copy --all "${SOURCE_REF}" "${TARGET_REF}"
-
+        MAX_RETRIES=3
+        RETRY_DELAY=10
+        for attempt in $(seq 1 $MAX_RETRIES); do
+          echo "Attempt ${attempt}/${MAX_RETRIES}..."
+          if skopeo copy --all "${SOURCE_REF}" "${TARGET_REF}"; then
             echo "target_image_ref=${{ inputs.target_registry }}/${TARGET_IMAGE}:${TARGET_TAG}" >> $GITHUB_OUTPUT
             echo "✅ Image copied successfully"
+            exit 0
+          fi
+          if [ "$attempt" -lt "$MAX_RETRIES" ]; then
+            echo "⚠️ Attempt ${attempt} failed, retrying in ${RETRY_DELAY}s..."
+            sleep $RETRY_DELAY
+            RETRY_DELAY=$((RETRY_DELAY * 2))
+          fi
+        done
+        echo "❌ All ${MAX_RETRIES} attempts failed"
+        exit 1

.github/workflows/build-test-distribute-flavor-matrix.yml

@@ -56,6 +56,26 @@ on:
         required: false
         type: boolean
         default: true
+      build_timeout_minutes:
+        description: 'Timeout in minutes for the build step'
+        required: false
+        type: number
+        default: 60
+      test_gpu_timeout_minutes:
+        description: 'Timeout in minutes for the GPU test step'
+        required: false
+        type: number
+        default: 30
+      test_cpu_timeout_minutes:
+        description: 'Timeout in minutes for the CPU test step'
+        required: false
+        type: number
+        default: 10
+      copy_timeout_minutes:
+        description: 'Timeout in minutes for the copy to ACR step'
+        required: false
+        type: number
+        default: 5
     secrets:
       AWS_DEFAULT_REGION:
         required: true
@@ -99,4 +119,8 @@ jobs:
       run_tests: ${{ inputs.run_tests && !(inputs.framework == 'trtllm' && matrix.platform == 'arm64') }} # trtllm tests on arm64 are not supported
       copy_to_acr: ${{ inputs.copy_to_acr && matrix.platform == 'amd64' }} # no reason to copy ARM images to ACR
       push_image: ${{ inputs.push_image }}
+      build_timeout_minutes: ${{ inputs.build_timeout_minutes }}
+      test_gpu_timeout_minutes: ${{ inputs.test_gpu_timeout_minutes }}
+      test_cpu_timeout_minutes: ${{ inputs.test_cpu_timeout_minutes }}
+      copy_timeout_minutes: ${{ inputs.copy_timeout_minutes }}
     secrets: inherit

.github/workflows/build-test-distribute-flavor.yml

@@ -66,6 +66,26 @@ on:
         required: false
         type: boolean
         default: false
+      build_timeout_minutes:
+        description: 'Timeout in minutes for the build step'
+        required: false
+        type: number
+        default: 60
+      test_gpu_timeout_minutes:
+        description: 'Timeout in minutes for the GPU test step'
+        required: false
+        type: number
+        default: 30
+      test_cpu_timeout_minutes:
+        description: 'Timeout in minutes for the CPU test step'
+        required: false
+        type: number
+        default: 10
+      copy_timeout_minutes:
+        description: 'Timeout in minutes for the copy to ACR step'
+        required: false
+        type: number
+        default: 5
     secrets:
       AWS_DEFAULT_REGION:
         required: true
@@ -185,6 +205,7 @@ jobs:
               --short-output
       - name: Build Container
         id: build-image
+        timeout-minutes: ${{ inputs.build_timeout_minutes }}
         uses: ./.github/actions/docker-remote-build
         with:
           image_tag: ${{ steps.calculate-target-tag.outputs.default_target_image_uri }}
@@ -275,6 +296,7 @@ jobs:
       # Run CPU-only tests first (parallelized for speed)
       # These are unit tests marked with gpu_0 that don't require GPU hardware
       - name: Run CPU-only tests (parallelized)
+        timeout-minutes: ${{ inputs.test_cpu_timeout_minutes }}
         uses: ./.github/actions/pytest
         with:
           image_tag: ${{ steps.calculate-target-tag.outputs.test_image }}
@@ -290,6 +312,7 @@ jobs:
       # Run GPU tests sequentially (only on amd64 runners with GPU)
       # These are e2e tests marked with gpu_1 that require GPU hardware
       - name: Run GPU tests (sequential)
+        timeout-minutes: ${{ inputs.test_gpu_timeout_minutes }}
         if: ${{ inputs.platform == 'amd64' }} # We only run GPU tests on amd64
         uses: ./.github/actions/pytest
         with:
@@ -333,6 +356,7 @@ jobs:
           echo "test_image=${TEST_IMAGE}" >> $GITHUB_OUTPUT
 
       - name: Copy image to target registry
+        timeout-minutes: ${{ inputs.copy_timeout_minutes }}
         uses: ./.github/actions/skopeo-copy
         with:
           source_registry: ${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com

.github/workflows/pr.yaml

@@ -184,6 +184,7 @@ jobs:
         ${{ github.ref_name == 'main' && 'main-vllm' || '' }}
         ${{ github.ref_name == 'main' && format('main-vllm-{0}', github.sha) || '' }}
       builder_name: ${{ needs.changed-files.outputs.builder_name }}
+      test_gpu_timeout_minutes: 35
     secrets: inherit
 
   # ============================================================================