Enable shared libraries.

.github/workflows/linux-cpu-arm64-build.yml

@@ -41,7 +41,7 @@ jobs:
             --docker-build-args "--build-arg BUILD_UID=$( id -u )" \
             --container-registry onnxruntimebuildcache \
             --repository onnxruntimecpubuild
-          docker run --rm --volume $GITHUB_WORKSPACE:/onnxruntime_src --volume $RUNNER_TEMP:/build -w /build onnxruntimecpubuild bash -c "echo $PATH && /usr/bin/cmake -G Ninja /onnxruntime_src -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release && ninja"
+          docker run --rm --volume $GITHUB_WORKSPACE:/onnxruntime_src --volume $RUNNER_TEMP:/build -w /build onnxruntimecpubuild bash -c "echo $PATH && /usr/bin/cmake -G Ninja /onnxruntime_src -DCMAKE_BUILD_TYPE=Release && ninja"
 
 
 

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

@@ -38,7 +38,7 @@ jobs:
           set -e -x
           mkdir build
           cd build
-          cmake -G "Unix Makefiles" .. -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
+          cmake -G "Unix Makefiles" .. -DCMAKE_BUILD_TYPE=Release
           cmake --build . --config Release --parallel
           cd ..
           rm -rf build
@@ -48,7 +48,7 @@ jobs:
           set -e -x
           mkdir build
           cd build
-          CC=clang CXX=clang++ cmake -G "Unix Makefiles" .. -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_CLANG_TIDY=clang-tidy
+          CC=clang CXX=clang++ cmake -G "Unix Makefiles" .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_CLANG_TIDY=clang-tidy
           cmake --build . --config Release --parallel
 
 

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

@@ -55,5 +55,5 @@ jobs:
             --rm \
             --volume $GITHUB_WORKSPACE:/onnxruntime_src \
             --volume $RUNNER_TEMP:/build \
-            -w /build onnxruntimegpubuild bash -c "/usr/bin/cmake -G Ninja /onnxruntime_src -DBUILD_SHARED_LIBS=ON -DCMAKE_POSITION_INDEPENDENT_CODE=ON -DCMAKE_CUDA_ARCHITECTURES=75 -DUSE_CUDA=1 -DCMAKE_BUILD_TYPE=Release && ninja"
+            -w /build onnxruntimegpubuild bash -c "/usr/bin/cmake -G Ninja /onnxruntime_src -DCMAKE_POSITION_INDEPENDENT_CODE=ON -DCMAKE_CUDA_ARCHITECTURES=75 -DUSE_CUDA=1 -DCMAKE_BUILD_TYPE=Release && ninja"
 

CMakeLists.txt

@@ -76,21 +76,20 @@ else()
   list(REMOVE_ITEM generator_srcs ${generator_cuda_srcs})
 endif()
 
-## TODO: this is a hack to get the shared library to export .lib files on windows. We should find a better way when we need to compile dynamic libraries
-#if(WIN32)
-#    set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
-#endif()
-
-add_library(onnxruntime-genai ${generator_srcs})
+add_library(onnxruntime-genai SHARED ${generator_srcs})
+add_library(onnxruntime-genai-static STATIC ${generator_srcs})
 target_include_directories(onnxruntime-genai PRIVATE ${CMAKE_SOURCE_DIR}/ort/include)
+target_include_directories(onnxruntime-genai-static PRIVATE ${CMAKE_SOURCE_DIR}/ort/include)
+
 
 if(USE_TOKENIZER)
   add_subdirectory("${CMAKE_SOURCE_DIR}/src/tokenizer")
   message("Using Tokenizer")
 endif()
 
 if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
-  set_target_properties(onnxruntime-genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
+  set_target_properties(onnxruntime-genai-static PROPERTIES POSITION_INDEPENDENT_CODE ON)
+
 endif()
 
 if(WIN32)
@@ -110,25 +109,28 @@ endif()
 add_executable(Tests ${test_srcs})
 target_include_directories(Tests PRIVATE ${CMAKE_SOURCE_DIR}/ort/include)
 target_link_directories(Tests PRIVATE ${CMAKE_SOURCE_DIR}/ort/lib)
-target_link_libraries(Tests PRIVATE onnxruntime-genai ${ONNXRUNTIME_LIB})
+target_link_libraries(Tests PRIVATE onnxruntime-genai-static ${ONNXRUNTIME_LIB})
 #
 pybind11_add_module(python ${python_srcs})
 target_include_directories(python PRIVATE ${CMAKE_SOURCE_DIR}/ort/include)
 target_link_directories(python PRIVATE ${CMAKE_SOURCE_DIR}/ort/lib)
-target_link_libraries(python PRIVATE onnxruntime-genai ${ONNXRUNTIME_LIB})
+target_link_libraries(python PRIVATE onnxruntime-genai-static ${ONNXRUNTIME_LIB})
 set_target_properties(python PROPERTIES OUTPUT_NAME "onnxruntime_genai")
 
-# TODO: Windows is statically linking CUDA libraries, so remove the 'NOT WIN32' once it's using shared libraries
-if(NOT WIN32 AND USE_CUDA AND CMAKE_CUDA_COMPILER)
+if(USE_CUDA AND CMAKE_CUDA_COMPILER)
   set_target_properties(onnxruntime-genai PROPERTIES LINKER_LANGUAGE CUDA)
-  set_target_properties(python PROPERTIES LINKER_LANGUAGE CUDA)
   set_target_properties(Tests PROPERTIES LINKER_LANGUAGE CUDA)
+  set_target_properties(python PROPERTIES LINKER_LANGUAGE CUDA)
   target_link_libraries(onnxruntime-genai PRIVATE cublasLt cublas cudnn curand cufft cudart)
   target_link_libraries(Tests PRIVATE cublasLt cublas cudnn curand cufft cudart)
   target_link_libraries(python PRIVATE cublasLt cublas cudnn curand cufft cudart)
+  #  onnxruntime-genai-static is statically linked under Windows
+  if(NOT WIN32)
+    set_target_properties(onnxruntime-genai-static PROPERTIES LINKER_LANGUAGE CUDA)
+    target_link_libraries(onnxruntime-genai-static PRIVATE cublasLt cublas cudnn curand cufft cudart)
+  endif()
 endif()
 
-
 # Visual C++ - /MP Multiple process build, /WX Warnings as errors
 # Enable:
 #   C5038: data member 'member1' will be initialized after data member 'member2'
@@ -142,34 +144,32 @@ endif()
 file(GLOB onnxruntime_libs "${CMAKE_SOURCE_DIR}/ort/${ONNXRUNTIME_FILES}")
 foreach(DLL_FILE ${onnxruntime_libs})
   add_custom_command(
-    TARGET onnxruntime-genai POST_BUILD
-    COMMAND ${CMAKE_COMMAND} -E copy ${DLL_FILE} ${CMAKE_BINARY_DIR}/$<CONFIG>/${DLL_FILE_NAME}
+     TARGET onnxruntime-genai POST_BUILD
+     COMMAND ${CMAKE_COMMAND} -E copy ${DLL_FILE} ${CMAKE_BINARY_DIR}/$<CONFIG>/${DLL_FILE_NAME}
   )
 endforeach()
 
 if(USE_TOKENIZER)
   add_compile_definitions(USE_TOKENIZER=1)
   target_include_directories(onnxruntime-genai PRIVATE ${TOKENIZER_ROOT})
+  target_include_directories(onnxruntime-genai-static PRIVATE ${TOKENIZER_ROOT})
   target_include_directories(Tests PRIVATE ${TOKENIZER_ROOT})
   target_include_directories(python PRIVATE ${TOKENIZER_ROOT})
-  target_link_libraries(Tests PRIVATE tokenizer)
-  target_link_libraries(python PRIVATE tokenizer)
+  target_link_libraries(onnxruntime-genai PRIVATE tokenizer)
+  target_link_libraries(onnxruntime-genai-static PRIVATE tokenizer)
   message("Linking Tokenizer with TOKENIZER_ROOT: ${TOKENIZER_ROOT}")
 endif()
 
 if(BUILD_WHEEL)
   set(WHEEL_FILES_DIR "${CMAKE_BINARY_DIR}/wheel")
   message("Setting up wheel files in : ${WHEEL_FILES_DIR}")
-
   set(TARGET_NAME "onnxruntime_genai")
-
   configure_file(${PYTHON_ROOT}/setup.py.in ${WHEEL_FILES_DIR}/setup.py @ONLY)
   configure_file(${PYTHON_ROOT}/py/__init__.py.in ${WHEEL_FILES_DIR}/${TARGET_NAME}/__init__.py @ONLY)
-
   file(GLOB onnxruntime_libs "${CMAKE_SOURCE_DIR}/ort/${ONNXRUNTIME_FILES}")
   foreach(DLL_FILE ${onnxruntime_libs})
     add_custom_command(
-       TARGET onnxruntime-genai
+       TARGET onnxruntime-genai-static
        COMMAND ${CMAKE_COMMAND} -E copy ${DLL_FILE} ${WHEEL_FILES_DIR}/${TARGET_NAME}/${DLL_FILE_NAME}
     )
   endforeach()

src/tokenizer/CMakeLists.txt

@@ -1,38 +1,43 @@
 set(TOKENIZER_ROOT ${CMAKE_CURRENT_SOURCE_DIR} PARENT_SCOPE)
 file(GLOB tokenizer_srcs CONFIGURE_DEPENDS
-        "${CMAKE_CURRENT_SOURCE_DIR}/*.cc"
-        "${CMAKE_CURRENT_SOURCE_DIR}/utils/*.cc"
-        "${CMAKE_CURRENT_SOURCE_DIR}/c_api/*.cc"
+   "${CMAKE_CURRENT_SOURCE_DIR}/*.cc"
+   "${CMAKE_CURRENT_SOURCE_DIR}/utils/*.cc"
+   "${CMAKE_CURRENT_SOURCE_DIR}/c_api/*.cc"
 )
 FetchContent_Declare(spm URL https://github.com/google/sentencepiece/archive/refs/tags/v0.1.99.zip)
 set(SPM_USE_EXTERNAL_ABSL OFF CACHE BOOL "Use external absl")
 set(SPM_USE_BUILTIN_PROTOBUF ON CACHE BOOL "Use built-in protobuf")
 set(SPM_ENABLE_SHARED OFF)
 
 FetchContent_GetProperties(spm)
-if (NOT spm_POPULATED)
-    FetchContent_Populate(spm)
-    add_subdirectory(${spm_SOURCE_DIR} ${spm_BINARY_DIR} EXCLUDE_FROM_ALL)
-endif ()
+if(NOT spm_POPULATED)
+  FetchContent_Populate(spm)
+  add_subdirectory(${spm_SOURCE_DIR} ${spm_BINARY_DIR} EXCLUDE_FROM_ALL)
+endif()
 
 FetchContent_Declare(GSL URL https://github.com/microsoft/GSL/archive/refs/tags/v4.0.0.zip)
 FetchContent_MakeAvailable(GSL)
 
 FetchContent_Declare(simdjson URL https://github.com/simdjson/simdjson/archive/refs/tags/v3.6.3.zip URL_HASH SHA1=2b063a2e81f74a5d1cb937fadf3d2fca0f1edb09)
 FetchContent_MakeAvailable(simdjson)
 
-add_library(tokenizer ${tokenizer_srcs})
+add_library(tokenizer STATIC ${tokenizer_srcs})
 message(STATUS "GSL_SOURCE_DIR: ${GSL_SOURCE_DIR}")
 target_include_directories(tokenizer PUBLIC
-        ${CMAKE_CURRENT_SOURCE_DIR}
-        ${GSL_SOURCE_DIR}/include
-        ${simdjson_SOURCE_DIR}/singleheader/
-        ${spm_SOURCE_DIR}/src
-        ${spm_SOURCE_DIR}/src/builtin_pb
-        ${spm_SOURCE_DIR}/third_party/absl
-        ${spm_SOURCE_DIR}/third_party/protobuf-lite
+   ${CMAKE_CURRENT_SOURCE_DIR}
+   ${GSL_SOURCE_DIR}/include
+   ${simdjson_SOURCE_DIR}/singleheader/
+   ${spm_SOURCE_DIR}/src
+   ${spm_SOURCE_DIR}/src/builtin_pb
+   ${spm_SOURCE_DIR}/third_party/absl
+   ${spm_SOURCE_DIR}/third_party/protobuf-lite
 )
 target_compile_definitions(tokenizer PRIVATE _SILENCE_ALL_CXX20_DEPRECATION_WARNINGS)
 
+if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
+  set_target_properties(tokenizer PROPERTIES POSITION_INDEPENDENT_CODE ON)
+  set_target_properties(simdjson PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif()
+
 add_dependencies(tokenizer sentencepiece-static)
 target_link_libraries(tokenizer PUBLIC sentencepiece-static simdjson)