Merge pull request #33 from microsoft/Cjian/fix-linux-gpu

Fix CUDAToolkit_ROOT

.github/workflows/linux-gpu-x64-build.yml

@@ -1,5 +1,5 @@
 name: "Linux GPU x64 Build"
-on: [ workflow_dispatch, pull_request ]
+on: [ workflow_dispatch ]
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}

.github/workflows/win-gpu-x64-build.yml

@@ -11,6 +11,8 @@ env:
   ort_dir: "onnxruntime-win-x64-gpu-1.16.3"
   ort_zip: "onnxruntime-win-x64-gpu-1.16.3.zip"
   ort_url: "https://github.com/microsoft/onnxruntime/releases/download/v1.16.3/onnxruntime-win-x64-gpu-1.16.3.zip"
+  cuda_dir: "${{ github.workspace }}\\cuda_sdk"
+  cuda_version: "12.2"
 
 jobs:
   Windows-CUDA-12-Build:
@@ -25,7 +27,8 @@ jobs:
           architecture: 'x64'
 
       - name: Download cuda
-        run: azcopy.exe cp --recursive "https://lotusscus.blob.core.windows.net/models/cuda_sdk/v12.2" cuda_sdk
+        run: |
+          azcopy.exe cp --recursive "https://lotusscus.blob.core.windows.net/models/cuda_sdk/v${{ env.cuda_version }}" ${{ env.cuda_dir}}
 
       - name: Download OnnxRuntime
         run: |
@@ -42,12 +45,12 @@ jobs:
       - name: Git Submodule Update
         run: |
           git submodule update --init --recursive
-      # The build machine doesn't have a GPU. So the value of CMAKE_CUDA_ARCHITECTURES doesn't matter.
+
       - name: Build with CMake
         run: |
           mkdir build
           cd build
-          cmake -G "Visual Studio 17 2022" -A x64 .. -DCMAKE_BUILD_TYPE=Release -DUSE_CUDA=TRUE 
+          cmake -G "Visual Studio 17 2022" -A x64 -T cuda=${{ env.cuda_dir }}\\v${{ env.cuda_version }} .. -DCMAKE_BUILD_TYPE=Release -DUSE_CUDA=TRUE
           cmake --build . --config Release --parallel
 
       - name: Verify Build Artifacts

CMakeLists.txt

@@ -13,7 +13,9 @@ set(USE_ORT_EXT 0) # "Build with Onnxruntime Extensions tokenizer support"
 include(CheckLanguage)
 check_language(CUDA)
 if (CMAKE_CUDA_COMPILER)
-    set (CUDA_FOUND TRUE)
+    message(STATUS "CUDA found CUDAToolkit_VERSION ${CUDAToolkit_VERSION} with CMAKE_CUDA_COMPILER ${CMAKE_CUDA_COMPILER}" )
+else()
+    message(STATUS "CUDA not found")
 endif()
 
 if("${CMAKE_C_COMPILER_ID}" STREQUAL "GNU" AND CMAKE_C_COMPILER_VERSION VERSION_LESS 9)
@@ -35,7 +37,7 @@ set(PYTHON_ROOT ${PROJECT_SOURCE_DIR}/src/python)
 
 # CUDA Being enabled will make it not a debug build without this option, so all of the C++ headers will complain
 # about a mismatch with the actual debug headers and it'll fail to link. I don't know why this happens, or if this is the best fix.
-if(USE_CUDA AND CUDA_FOUND AND CMAKE_BUILD_TYPE STREQUAL "Debug")
+if(USE_CUDA AND CMAKE_CUDA_COMPILER AND CMAKE_BUILD_TYPE STREQUAL "Debug")
   add_compile_definitions(_DEBUG=1)
 endif()
 
@@ -58,7 +60,7 @@ file(GLOB python_srcs CMAKE_CONFIGURE_DEPENDS
    "${PYTHON_ROOT}/*.cpp"
 )
 
-if(USE_CUDA AND CUDA_FOUND)
+if(USE_CUDA AND CMAKE_CUDA_COMPILER)
     # Don't let cmake set a default value for CMAKE_CUDA_ARCHITECTURES
     cmake_policy(SET CMP0104 OLD)
     enable_language(CUDA)