Run libtorch examples, export error "undefined reference to xxx"

[TheGreatGalaxy]

🐛 Describe the bug

When run the pytorch example: https://pytorch.org/cppdocs/installing.html

CMakeLists.txt files:

cmake_minimum_required(VERSION 3.0 FATAL_ERROR)
project(example-app)

find_package(Torch REQUIRED)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${TORCH_CXX_FLAGS}")
# message("TORCH_CXX_FLAGS====: " ${TORCH_CXX_FLAGS})

# link_directories("/usr/local/cuda-11.2/targets/x86_64-linux/lib")

add_executable(example-app example-app.cpp)
target_link_libraries(example-app "${TORCH_LIBRARIES}")
message("TORCH_LIBRARIES:" "${TORCH_LIBRARIES}")
set_property(TARGET example-app PROPERTY CXX_STANDARD 14)

if (MSVC)
  file(GLOB TORCH_DLLS "${TORCH_INSTALL_PREFIX}/lib/*.dll")
  add_custom_command(TARGET example-app
                     POST_BUILD
                     COMMAND ${CMAKE_COMMAND} -E copy_if_different
                     ${TORCH_DLLS}
                     $<TARGET_FILE_DIR:example-app>)
endif (MSVC)

cpp:

#include <torch/torch.h>
#include <iostream>

int main() {
  torch::Tensor tensor = torch::rand({2, 3});
  std::cout << tensor << std::endl;
}

cmake command:

cmake -DCMAKE_PREFIX_PATH=/home/guangtong/project/libtorch  ..

cmake --build . --config Release

will appears:

cmake --build . --config Release
[ 50%] Linking CXX executable example-app
/home/xxxxxx/project/libtorch/lib/libtorch_cuda_cpp.so: undefined reference to `cudaStreamUpdateCaptureDependencies@libcudart.so.11.0'
/home/xxxxxx/project/libtorch/lib/libtorch_cuda_cpp.so: undefined reference to `cudaStreamGetCaptureInfo_v2@libcudart.so.11.0'
/home/xxxxxx/project/libtorch/lib/libtorch_cuda_cpp.so: undefined reference to `cudaGraphRetainUserObject@libcudart.so.11.0'
/home/xxxxxx/project/libtorch/lib/libtorch_cuda_cpp.so: undefined reference to `cudaUserObjectCreate@libcudart.so.11.0'
collect2: error: ld returned 1 exit status
CMakeFiles/example-app.dir/build.make:110: recipe for target 'example-app' failed
make[2]: *** [example-app] Error 1
CMakeFiles/Makefile2:67: recipe for target 'CMakeFiles/example-app.dir/all' failed
make[1]: *** [CMakeFiles/example-app.dir/all] Error 2
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2

but seems the file dependencies is ok：

ldd -r  libtorch_cuda_cpp.so
	linux-vdso.so.1 (0x00007ffe39818000)
	libcudart-a7b20f20.so.11.0 (0x00007fba89fc1000)
	libc10_cuda.so (0x00007fba89d5b000)
	libnvToolsExt-24de1d56.so.1 (0x00007fba89b51000)
	libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007fba89932000)
	libc10.so (0x00007fba896ba000)
	libtorch_cpu.so (0x00007fba726f1000)
	libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007fba724ed000)
	librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007fba722e5000)
	libstdc++.so.6 => /usr/lib/x86_64-linux-gnu/libstdc++.so.6 (0x00007fba71f5c000)
	libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007fba71bbe000)
	libgomp-52f2fd74.so.1 (0x00007fba7198b000)
	libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007fba71773000)
	libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007fba71382000)
	/lib64/ld-linux-x86-64.so.2 (0x00007fbafb5a3000)

Versions

libtorch version:
https://download.pytorch.org/libtorch/cu113/libtorch-cxx11-abi-shared-with-deps-1.10.2%2Bcu113.zip

os: Ubuntu 18.04

cmake version: 3.10.2

cc @ezyang @seemethere @malfet

[adeilsonsilva]

Check your CUDA version using nvidia-smi, you probably have CUDA < 11.3. I was trying to use the same libtorch version as you with CUDA 11.2 and had the same errors. After I downloaded libtorch for cuda 11.1 I was able to compile succesfully.

[TheGreatGalaxy]

Check your CUDA version using nvidia-smi, you probably have CUDA < 11.3. I was trying to use the same libtorch version as you with CUDA 11.2 and had the same errors. After I downloaded libtorch for cuda 11.1 I was able to compile succesfully.

My cuda is 11.4. I will try with cuda 11.1. Thanks!

[waldirfc]

I still had the same problem with libtorch cu113 and cuda 11.4.
Meanwhile, I am working on the CPU version.
If you have any news I would appreciate them.

[Fatichti]

I try to install pytorch from source.
i have CUDA 11.7 on my computer.

i have build successfully but i have an error when i try to import a module from d2go

Error initial :
Failed to load image Python extension: /home/kainoo/miniconda3/envs/Env4D2go/lib/python3.9/site-packages/torch/lib/libtorch_cuda_cpp.so: undefined symbol: _ZN5torch19kParamCommsCallNameE

Error detail :

Link lib detail:

Do you have any idea to solve it ?

[MrRace]

nvcc --version

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Wed_Jun__8_16:49:14_PDT_2022
Cuda compilation tools, release 11.7, V11.7.99
Build cuda_11.7.r11.7/compiler.31442593_0

Therefore I choose the https://download.pytorch.org/libtorch/cu117/libtorch-cxx11-abi-shared-with-deps-1.13.0%2Bcu117.zip
And follow the guideline in https://pytorch.org/cppdocs/installing.html and when run the ./example-app, but error comes:

./example-app: symbol lookup error: /share/tools/libtorch-cxx11-abi-shared-with-deps-1.13.0+cu117/lib/libtorch_cuda_cpp.so: undefined symbol: _ZTIN4c10d4WorkE

How to solve it ? Thanks a lot!

[Wentao795]

nvcc --version

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Wed_Jun__8_16:49:14_PDT_2022
Cuda compilation tools, release 11.7, V11.7.99
Build cuda_11.7.r11.7/compiler.31442593_0

因此我选择了https://download.pytorch.org/libtorch/cu117/libtorch-cxx11-abi-shared-with-deps-1.13.0%2Bcu117.zip And follow the guideline inhttps://pytorch.org/cppdocs/installing.html和 when run the ./example-app，但是错误来了：

./example-app: symbol lookup error: /share/tools/libtorch-cxx11-abi-shared-with-deps-1.13.0+cu117/lib/libtorch_cuda_cpp.so: undefined symbol: _ZTIN4c10d4WorkE

如何解决？非常感谢！

hi ~, i have the same problem,you have solved it？

[thiagocrepaldi]

same here, but the error only happens when I try to use torch::randn({1, 10, 12, 12}). If I use other torch APIs like torch::relu, there is no linking issue.

[puppy2333]

I think using matched cuda and libtorch versions solves this issue. Moreover maybe we should only use the listed versions of libtorch on the official website (currently 11.8, 12.1 and 12.4)?