编译PyTorch转ONNX模型时CMake无法找到ONNX Runtime库求助

编译C++ ONNX推理程序时遇到以下错误：

CMake Error : The following variables are used in this project, but they are set to NOTFOUND. Please set them or make sure they are set and tested correctly in the CMake files : ONNX_RUNTIME_LIB liked by target "ObjectDetectionInferende" in directory /hyeonseop/OD/test/onnx_od

当前使用的CMake代码：

cmake_minimum_required(VERSION 3.25)
project(ObjectDetectionInference)

find_package(OpenCV REQUIRED)

find_path(ONNX_RUNTIME_SESSION_INCLUDE_DIRS onnxruntime_cxx_api.h HINTS "/hyeonseop/OD/onnxruntime-linux-x64-gpu-1.17.0/include")
find_library(ONNX_RUNTIME_LIB lib HINTS "/hyeonseop/OD/onnxruntime-linux-x64-gpu-1.17.0")

set(SOURCE_FILES main.cpp helper.cpp)

add_executable(ObjectDetectionInference ${SOURCE_FILES})

target_include_directories(ObjectDetectionInference PRIVATE ${ONNX_RUNTIME_SESSION_INCLUDE_DIRS}) # ${OpenCV_INCLUDE_DIRS}
target_link_libraries(ObjectDetectionInference PRIVATE ${ONNX_RUNTIME_LIB}) # ${OpenCV_LIBRARIES}

ONNX Runtime安装路径为/hyeonseop/OD/onnxruntime-linux-x64-gpu-1.17.0，目录结构：

root@4fe7a1a9f82d:/hyeonseop/OD/onnxruntime-linux-x64-gpu-1.17.0# ls
GIT_COMMIT_ID LICENSE Privacy.md README.md ThirdPartyNotices.txt
VERSION_NUMBER include lib

问题根源 #

find_library的第二个参数传错了——它需要的是库文件的基础名称，而不是目录名lib。你当前的写法会让CMake去寻找名为lib的库，显然不存在。

修复方案 #

1. 修正find_library调用 #

将find_library改为指向lib子目录，并指定正确的库名称：

find_library(ONNX_RUNTIME_LIB 
             NAMES onnxruntime_gpu onnxruntime  # 优先匹配GPU版本库
             HINTS "/hyeonseop/OD/onnxruntime-linux-x64-gpu-1.17.0/lib")

• NAMES后填库的基础名：Linux下库文件是libonnxruntime_gpu.so，CMake会自动补全lib前缀和.so后缀。
• HINTS直接指向lib子目录，缩小搜索范围。

2. 添加存在性检查（可选但推荐） #

在find_path和find_library后添加检查，提前暴露问题：

if(NOT ONNX_RUNTIME_SESSION_INCLUDE_DIRS)
    message(FATAL_ERROR "未找到ONNX Runtime头文件目录")
endif()
if(NOT ONNX_RUNTIME_LIB)
    message(FATAL_ERROR "未找到ONNX Runtime库文件")
endif()

3. 恢复OpenCV依赖链接 #

如果你的代码用到了OpenCV，必须把注释掉的OpenCV依赖恢复：

target_include_directories(ObjectDetectionInference PRIVATE 
    ${ONNX_RUNTIME_SESSION_INCLUDE_DIRS} 
    ${OpenCV_INCLUDE_DIRS})
target_link_libraries(ObjectDetectionInference PRIVATE 
    ${ONNX_RUNTIME_LIB} 
    ${OpenCV_LIBRARIES})

额外说明 #

• 不需要重新安装ONNX Runtime，路径配置本身没问题，只是CMake调用参数错误。
• 确认/hyeonseop/OD/onnxruntime-linux-x64-gpu-1.17.0/lib下存在libonnxruntime_gpu.so或libonnxruntime.so文件，若只有带版本号的文件（如libonnxruntime.so.1.17.0），CMake也能自动识别链接。

内容的提问来源于stack exchange，提问作者Hyeonseop Lim