Add CUDA Plugin EP CI and fix Windows plugin build support

.github/workflows/linux_cuda_plugin_ci.yml

@@ -0,0 +1,126 @@
+name: CUDA Plugin Linux CI
+
+on:
+  push:
+    branches: [main, 'rel-*']
+  pull_request:
+    branches: [main, 'rel-*']
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event_name == 'pull_request' && github.ref || github.sha }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+  packages: write
+  attestations: write
+  id-token: write
+
+jobs:
+  build-linux-cuda-plugin-x64-release:
+    name: Build Linux CUDA Plugin EP x64 Release
+    uses: ./.github/workflows/reusable_linux_build.yml
+    with:
+      pool_name: "onnxruntime-github-Ubuntu2204-AMD-CPU"
+      build_config: Release
+      architecture: x64
+      dockerfile_path: tools/ci_build/github/linux/docker/Dockerfile.manylinux2_28_cuda
+      docker_build_args: '--build-arg BASEIMAGE=onnxruntimebuildcache.azurecr.io/internal/azureml/onnxruntime/build/cuda12_x64_almalinux8_gcc14:20251017.1'
+      docker_image_repo: onnxruntimecuda12manylinuxbuild
+      extra_build_flags: >-
+        --use_binskim_compliant_compile_flags
+        --build_wheel
+        --parallel
+        --nvcc_threads 1
+        --cuda_version=12.8
+        --cuda_home=/usr/local/cuda-12.8
+        --cudnn_home=/usr/local/cuda-12.8
+        --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86
+        --cmake_extra_defines onnxruntime_BUILD_CUDA_EP_AS_PLUGIN=ON
+      python_path_prefix: 'PATH=/opt/python/cp312-cp312/bin:$PATH'
+      run_tests: false
+      upload_build_output: true
+      execution_providers: 'cuda'
+      job_identifier: build-linux-cuda-plugin-x64-release
+    secrets:
+      GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+  test-linux-cuda-plugin-x64-release:
+    name: Test Linux CUDA Plugin EP x64 Release
+    needs: build-linux-cuda-plugin-x64-release
+    runs-on:
+      - self-hosted
+      - "1ES.Pool=onnxruntime-github-linux-a10"
+      - "JobId=test-linux-cuda-plugin-x64-release-${{ github.run_id }}-${{ github.run_number }}-${{ github.run_attempt }}"
+    permissions:
+      contents: read
+      packages: read
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v6
+
+      - uses: microsoft/onnxruntime-github-actions/build-docker-image@8bad63a3c05d448311dfa8e5f531171c97471aa1 # v0.0.12
+        id: build_docker_image_step
+        with:
+          dockerfile: ${{ github.workspace }}/tools/ci_build/github/linux/docker/Dockerfile.manylinux2_28_cuda
+          image-name: ghcr.io/microsoft/onnxruntime/onnxruntimecuda12manylinuxbuild
+          build-args: '--build-arg BASEIMAGE=onnxruntimebuildcache.azurecr.io/internal/azureml/onnxruntime/build/cuda12_x64_almalinux8_gcc14:20251017.1'
+          push: true
+          azure-container-registry-name: onnxruntimebuildcache
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+      # --- Download Build Artifact to Runner Temp Directory ---
+      - name: Download Build Artifact
+        uses: actions/download-artifact@v7
+        with:
+          name: build-output-x64-Release
+          path: ${{ runner.temp }}/Release
+
+      # --- Restore Permissions in the Temp Directory ---
+      - name: Restore Executable Permissions
+        if: success()
+        working-directory: ${{ runner.temp }}/Release
+        run: |
+          if [ -f perms.txt ]; then
+            echo "Restoring executable permissions in ${{ runner.temp }}/Release ..."
+            while IFS= read -r file; do
+              if [ -f "$file" ]; then
+                chmod +x "$file"
+              else
+                echo "Warning: File '$file' listed in perms.txt not found."
+              fi
+            done < perms.txt
+            echo "Permissions restored."
+          else
+            echo "Warning: perms.txt not found in artifact."
+          fi
+
+      # --- Install the ORT wheel and run CUDA plugin EP tests ---
+      - name: Run CUDA Plugin EP Python Tests
+        run: |
+          docker run --rm --gpus all \
+            -v ${{ github.workspace }}:/onnxruntime_src \
+            -v ${{ runner.temp }}/Release:/build/Release \
+            -e NVIDIA_VISIBLE_DEVICES=all \
+            ${{ steps.build_docker_image_step.outputs.full-image-name }} \
+            bash -c "
+              set -ex
+              export PATH=/opt/python/cp312-cp312/bin:\$PATH
+
+              # Install the ORT wheel
+              python -m pip install /build/Release/Release/dist/onnxruntime*.whl
+
+              # Install test dependencies
+              python -m pip install numpy onnx
+              python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
+
+              # Set plugin path and run tests
+              export ORT_CUDA_PLUGIN_PATH=/build/Release/Release/libonnxruntime_providers_cuda_plugin.so
+              echo \"ORT_CUDA_PLUGIN_PATH=\$ORT_CUDA_PLUGIN_PATH\"
+              ls -la \$ORT_CUDA_PLUGIN_PATH
+
+              cd /onnxruntime_src/onnxruntime/test/python/transformers
+              python test_cuda_plugin_ep.py
+            "

.github/workflows/windows_cuda_plugin.yml

@@ -0,0 +1,207 @@
+name: CUDA Plugin Windows CI
+
+on:
+  push:
+    branches:
+      - main
+      - rel-*
+  pull_request:
+    branches:
+      - main
+      - rel-*
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event_name == 'pull_request' && github.ref || github.sha }}
+  cancel-in-progress: true
+
+jobs:
+  build:
+    name: Windows CUDA Plugin EP Build
+    runs-on: [
+        "self-hosted",
+        "1ES.Pool=onnxruntime-github-vs2022-latest",
+        "JobId=windows-cuda-plugin-${{ github.run_id }}-${{ github.run_number }}-${{ github.run_attempt }}"
+        ]
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          submodules: 'none'
+
+      - uses: actions/setup-python@v6
+        with:
+          python-version: '3.14'
+          architecture: x64
+
+      - name: Locate vcvarsall and Setup Env
+        uses: ./.github/actions/locate-vcvarsall-and-setup-env
+        with:
+          architecture: x64
+
+      - name: Install python modules
+        run: python -m pip install -r .\tools\ci_build\github\windows\python\requirements.txt
+        working-directory: ${{ github.workspace }}
+        shell: cmd
+
+      - name: Download CUDA SDK v12.8
+        working-directory: ${{ runner.temp }}
+        run: |
+          azcopy.exe cp --recursive "https://lotusscus.blob.core.windows.net/models/cuda_sdk/v12.8" .
+          dir
+        shell: pwsh
+
+      - name: Add CUDA to PATH
+        shell: powershell
+        run: |
+          Write-Host "Adding CUDA to PATH"
+          Write-Host "CUDA Path: $env:RUNNER_TEMP\v12.8\bin"
+          Add-Content -Path $env:GITHUB_PATH -Value "$env:RUNNER_TEMP\v12.8\bin"
+          Add-Content -Path $env:GITHUB_PATH -Value "$env:RUNNER_TEMP\v12.8\extras\CUPTI\lib64"
+
+      - name: Set OnnxRuntimeBuildDirectory
+        shell: pwsh
+        run: |
+          $buildDir = Join-Path ${{ runner.temp }} "build"
+          echo "OnnxRuntimeBuildDirectory=$buildDir" >> $env:GITHUB_ENV
+
+      - name: Build ONNX Runtime with CUDA Plugin EP
+        working-directory: ${{ runner.temp }}
+        run: |
+          python.exe ${{ github.workspace }}\tools\ci_build\build.py `
+            --update --build --config Release `
+            --build_dir build `
+            --skip_submodule_sync `
+            --parallel `
+            --nvcc_threads 1 `
+            --use_binskim_compliant_compile_flags `
+            --cmake_generator "Visual Studio 17 2022" `
+            --build_shared_lib `
+            --build_wheel `
+            --use_cuda `
+            --cuda_home="$env:RUNNER_TEMP\v12.8" `
+            --skip_tests `
+            --use_vcpkg `
+            --use_vcpkg_ms_internal_asset_cache `
+            --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86 `
+            --cmake_extra_defines onnxruntime_BUILD_CUDA_EP_AS_PLUGIN=ON
+
+          if ($lastExitCode -ne 0) {
+            exit $lastExitCode
+          }
+
+          # Clean up intermediate files before uploading artifacts
+          $outputDir = "${{ runner.temp }}\build\Release"
+          Write-Host "Cleaning up files from $outputDir..."
+
+          Remove-Item -Path "$outputDir\onnxruntime" -Recurse -Force -ErrorAction SilentlyContinue
+          Remove-Item -Path "$outputDir\pybind11" -Recurse -Force -ErrorAction SilentlyContinue
+          Remove-Item -Path "$outputDir\models" -Recurse -Force -ErrorAction SilentlyContinue
+          Remove-Item -Path "$outputDir\vcpkg_installed" -Recurse -Force -ErrorAction SilentlyContinue
+          Remove-Item -Path "$outputDir\_deps" -Recurse -Force -ErrorAction SilentlyContinue
+          Remove-Item -Path "$outputDir\CMakeCache.txt" -Force -ErrorAction SilentlyContinue
+          Remove-Item -Path "$outputDir\CMakeFiles" -Recurse -Force -ErrorAction SilentlyContinue
+          Remove-Item -Path $outputDir -Include "*.obj" -Recurse
+        shell: pwsh
+
+      - name: Upload build artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          name: cuda-plugin-build-artifacts
+          path: ${{ runner.temp }}\build
+    env:
+      DOTNET_SKIP_FIRST_TIME_EXPERIENCE: true
+      setVcvars: true
+      ALLOW_RELEASED_ONNX_OPSET_ONLY: '0'
+      ONNXRUNTIME_TEST_GPU_DEVICE_ID: '0'
+      AZCOPY_AUTO_LOGIN_TYPE: MSI
+      AZCOPY_MSI_CLIENT_ID: 63b63039-6328-442f-954b-5a64d124e5b4
+
+  test:
+    name: Windows CUDA Plugin EP Test
+    needs: build
+    timeout-minutes: 120
+    runs-on: [
+      "self-hosted",
+      "1ES.Pool=onnxruntime-github-Win2022-GPU-A10",
+      "JobId=windows-cuda-plugin-test-${{ github.run_id }}-${{ github.run_number }}-${{ github.run_attempt }}"
+    ]
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          submodules: 'none'
+
+      - name: Download build artifacts
+        uses: actions/download-artifact@v7
+        with:
+          name: cuda-plugin-build-artifacts
+          path: ${{ runner.temp }}\build
+
+      - uses: actions/setup-python@v6
+        with:
+          python-version: '3.14'
+          architecture: x64
+
+      - name: Locate vcvarsall and Setup Env
+        uses: ./.github/actions/locate-vcvarsall-and-setup-env
+        with:
+          architecture: x64
+
+      - name: Install python modules
+        run: python -m pip install -r .\tools\ci_build\github\windows\python\requirements.txt
+        working-directory: ${{ github.workspace }}
+        shell: cmd
+
+      - name: Install torch for CPU only
+        run: python -m pip install torch
+        working-directory: ${{ github.workspace }}
+        shell: cmd
+
+      - name: Download CUDA SDK v12.8
+        working-directory: ${{ runner.temp }}
+        run: |
+          azcopy.exe cp --recursive "https://lotusscus.blob.core.windows.net/models/cuda_sdk/v12.8" .
+          dir
+        shell: pwsh
+
+      - name: Add CUDA to PATH
+        shell: powershell
+        run: |
+          Write-Host "Adding CUDA to PATH"
+          Write-Host "CUDA Path: $env:RUNNER_TEMP\v12.8\bin"
+          Add-Content -Path $env:GITHUB_PATH -Value "$env:RUNNER_TEMP\v12.8\bin"
+          Add-Content -Path $env:GITHUB_PATH -Value "$env:RUNNER_TEMP\v12.8\extras\CUPTI\lib64"
+
+      - name: Set OnnxRuntimeBuildDirectory
+        shell: pwsh
+        run: |
+          $buildDir = Join-Path ${{ runner.temp }} "build"
+          echo "OnnxRuntimeBuildDirectory=$buildDir" >> $env:GITHUB_ENV
+
+      - name: Install ONNX Runtime Wheel
+        uses: ./.github/actions/install-onnxruntime-wheel
+        with:
+          whl-directory: ${{ runner.temp }}\build\Release\Release\dist
+
+      - name: Run CUDA Plugin EP Python Tests
+        working-directory: ${{ github.workspace }}\onnxruntime\test\python\transformers
+        shell: pwsh
+        run: |
+          $env:ORT_CUDA_PLUGIN_PATH = "${{ runner.temp }}\build\Release\Release\onnxruntime_providers_cuda_plugin.dll"
+          Write-Host "ORT_CUDA_PLUGIN_PATH=$env:ORT_CUDA_PLUGIN_PATH"
+          if (-not (Test-Path $env:ORT_CUDA_PLUGIN_PATH)) {
+            Write-Error "CUDA plugin EP library not found at $env:ORT_CUDA_PLUGIN_PATH"
+            exit 1
+          }
+          python test_cuda_plugin_ep.py
+          if ($lastExitCode -ne 0) {
+            exit $lastExitCode
+          }
+    env:
+      DOTNET_SKIP_FIRST_TIME_EXPERIENCE: true
+      setVcvars: true
+      ALLOW_RELEASED_ONNX_OPSET_ONLY: '0'
+      ONNXRUNTIME_TEST_GPU_DEVICE_ID: '0'
+      AZCOPY_AUTO_LOGIN_TYPE: MSI
+      AZCOPY_MSI_CLIENT_ID: 63b63039-6328-442f-954b-5a64d124e5b4

cmake/onnxruntime_providers_cuda_plugin.cmake

@@ -143,10 +143,22 @@ target_compile_options(onnxruntime_providers_cuda_plugin PRIVATE
     "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:--std c++20>"
     "$<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr;-Xcudafe;--diag_suppress=550>"
     "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:-Xcudafe --diag_suppress=2810>"
-    "$<$<COMPILE_LANGUAGE:CXX>:-include;${REPO_ROOT}/include/onnxruntime/ep/adapters.h>"
-    "$<$<COMPILE_LANGUAGE:CXX>:SHELL:-include ${CUDA_PLUGIN_EP_DIR}/cuda_kernel_adapter.h>"
 )
 
+# Force-include adapter headers for CXX files.
+# MSVC uses /FI; GCC/Clang use -include.
+if (MSVC)
+    target_compile_options(onnxruntime_providers_cuda_plugin PRIVATE
+        "$<$<COMPILE_LANGUAGE:CXX>:SHELL:/FI \"${REPO_ROOT}/include/onnxruntime/ep/adapters.h\">"
+        "$<$<COMPILE_LANGUAGE:CXX>:SHELL:/FI \"${CUDA_PLUGIN_EP_DIR}/cuda_kernel_adapter.h\">"
+    )
+else()
+    target_compile_options(onnxruntime_providers_cuda_plugin PRIVATE
+        "$<$<COMPILE_LANGUAGE:CXX>:SHELL:-include ${REPO_ROOT}/include/onnxruntime/ep/adapters.h>"
+        "$<$<COMPILE_LANGUAGE:CXX>:SHELL:-include ${CUDA_PLUGIN_EP_DIR}/cuda_kernel_adapter.h>"
+    )
+endif()
+
 if (MSVC)
     target_compile_options(onnxruntime_providers_cuda_plugin PRIVATE
         "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:-Xcompiler /permissive>"