Unable to build and use libtorch function via pybind11: undefined symbol error upon import #73016

anubane · 2022-02-17T17:11:11Z

🐛 Describe the bug

I am attempting to write a C++ function in libtorch, compile it with pybind11 and then import the .so file in a python file.

The build seems to work fine (no errors are emitted) but, when I try to import it into my python file, execution fails with the error message:
ImportError: /home/cppNpython/MyProj/cpp/build/srcfile.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZN3c106detail23torchInternalAssertFailEPKcS2_jS2_RKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE

This is my directory structure:

|-- MyProj
|   |-- cpp
|   |   |-- build
|   |   |   `-- __init__.py
|   |   |-- include
|   |   |   `-- srcfile.h
|   |   |-- lib
|   |   |   |-- libtorch
|   |   |   `-- pybind11
|   |   |-- source
|   |   |   `-- srcfile.cpp
|   |   `-- CMakeLists.txt
|   `-- __init__.py
`-- test.py

Here are the contents of each file:-
srcfile.cpp

#include <srcfile.h>

torch::Tensor test(torch::Tensor t1, torch::Tensor t2) {
  return t1;
}

PYBIND11_MODULE(srcfile, m) {
    m.doc() = "demo function"; // optional module docstring
    m.def("test", &test, "A function giving torch tensor");
}

srcfile.h

#ifndef __SRCFILE_H__
#define __SRCFILE_H__

#include <torch/torch.h>
#include <pybind11/pybind11.h>

torch::Tensor test(torch::Tensor, torch::Tensor);

#endif

CMakeLists.txt

project(srcfile)
cmake_minimum_required(VERSION 3.10)
set(CMAKE_CXX_STANDARD 14)

find_package(PythonLibs REQUIRED)
set(CMAKE_PREFIX_PATH "lib/libtorch/share/cmake/Torch")
find_package(Torch REQUIRED)

include_directories(lib/libtorch)

include_directories(${PYTHON_INCLUDE_DIRS})

add_subdirectory(lib/pybind11)
pybind11_add_module(srcfile src/srcfile.cpp)

target_include_directories(srcfile PRIVATE include/)
target_link_libraries(srcfile PUBLIC ${TORCH_LIBRARIES})

set_property(TARGET srcfile PROPERTY CXX_STANDARD 14)

and at last test.py

import torch
from MyProj.cpp.build import srcfile as j

def main():
    t1 = torch.rand(2,2)
    t2 = torch.rand(2,2)
    print(j.test(t1, t2))


if __name__ == '__main__':
    main()

Inside the build directory I am executing the following command at the terminal:

cmake .. \
-DPYTHON_INCLUDE_DIR=$(python -c "from distutils.sysconfig import get_python_inc; print(get_python_inc())")  \
-DPYTHON_LIBRARY=$(python -c "import distutils.sysconfig as sysconfig; print(sysconfig.get_config_var('LIBDIR'))") \
-DTORCH_LIBRARIES=lib/libtorch && make

which does not emit any error.
I get the error only when I am attempting to import.

I am not sure what I am doing wrong here. Any help/suggestion is welcome.

Versions

PyTorch version: 1.10.2
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: Ubuntu 18.04.5 LTS (x86_64)
GCC version: (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
Clang version: Could not collect
CMake version: version 3.10.2
Libc version: glibc-2.9

Python version: 3.7.0 (default, Oct  9 2018, 10:31:47)  [GCC 7.3.0] (64-bit runtime)
Python platform: Linux-4.15.0-162-generic-x86_64-with-debian-buster-sid
Is CUDA available: False
CUDA runtime version: No CUDA
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A

Versions of relevant libraries:
[pip3] numpy==1.21.5
[pip3] torch==1.10.2
[conda] blas                      1.0                         mkl  
[conda] cpuonly                   2.0                           0    pytorch
[conda] mkl                       2022.0.1           h06a4308_117  
[conda] numpy                     1.21.5                   pypi_0    pypi
[conda] pytorch                   1.10.2              py3.7_cpu_0    pytorch
[conda] pytorch-mutex             1.0                         cpu    pytorch

cc @jbschlosser @gujinghui @PenghuiCheng @XiaobingSuper @jianyuh @VitalyFedyunin

The text was updated successfully, but these errors were encountered:

anubane · 2022-02-17T18:48:25Z

The error message:
undefined symbol: _ZN3c106detail23torchInternalAssertFailEPKcS2_jS2_RKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE

Upon demangling looks as follows:
undefined symbol: c10::detail::torchInternalAssertFail(char const*, char const*, unsigned int, char const*, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&)

Any thoughts on what might be wrong here?

anubane · 2022-02-18T08:52:18Z

I was previously using the cxx11 ABI version of the libtorch package available on pytorch.org.

Just for an experiment, I used the Pre-cxx11 ABI version and now I am getting the following error:

Traceback (most recent call last):
  File "../../../test.py", line 14, in <module>
    main()
  File "../../../test.py", line 10, in main
    print(j.test(t1, t2))
TypeError: test(): incompatible function arguments. The following argument types are supported:
    1. (arg0: at::Tensor, arg1: at::Tensor) -> at::Tensor

Invoked with: tensor([[0.4624, 0.3462],
        [0.4250, 0.3891]]), tensor([[0.6797, 0.2709],
        [0.5302,  0.7420]])

Is this error related to #20356?

anubane · 2022-03-17T11:19:35Z

I guess currently cannot do this via CMAKE, but only via setuptools, which PyTorch has already included via cpp_extensions

I am able to build the project now but cannot debug libtorch.

Issue at vscode-cpptools

ziyuang · 2022-04-01T17:33:18Z

I have a similar issue where _Z16THPVariable_WrapN2at10TensorBaseE (or THPVariable_Wrap(at::TensorBase)) becomes undefined. I have tried both cxx11 ABI and pre-cxx11 ABI but neither works. By "via setuptools" do you mean this part in the official documentation? https://pytorch.org/tutorials/advanced/cpp_extension.html#building-with-setuptools

roastduck · 2022-06-23T06:27:05Z

Any update on this thread? Same issue here of missing _ZN3c106detail23torchInternalAssertFailEPKcS2_jS2_RKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE. In my case, the issue only occurs when building with -DCMAKE_BUILD_TYPE=Debug. Maybe we have to build the C++ extension in the same CMAKE_BUILD_TYPE with the one used to build PyTorch itself?

roastduck · 2022-06-23T07:29:18Z

Defining a macro named STRIP_ERROR_MESSAGES solves my problem. Maybe PyTorch defines this macro in non-debug mode.

Jay-Ye · 2022-08-07T06:49:03Z

I was previously using the cxx11 ABI version of the libtorch package available on pytorch.org.

Just for an experiment, I used the Pre-cxx11 ABI version and now I am getting the following error:
Traceback (most recent call last):
  File "../../../test.py", line 14, in <module>
    main()
  File "../../../test.py", line 10, in main
    print(j.test(t1, t2))
TypeError: test(): incompatible function arguments. The following argument types are supported:
    1. (arg0: at::Tensor, arg1: at::Tensor) -> at::Tensor

Invoked with: tensor([[0.4624, 0.3462],
        [0.4250, 0.3891]]), tensor([[0.6797, 0.2709],
        [0.5302,  0.7420]])
Is this error related to #20356?

@anubane
I'm getting the exact same error like this, "incompatible function arguments". Did you solve this by compiling the code via setup.py with CppExtension and BuildExtension in torch.utils.cpp_extension?

mananshah99 · 2022-09-16T06:54:46Z

Running into the same problem as well, even when building torch using the same libtorch.so as my python torch installation. Any chance there is a recommended solution?

anubane · 2022-09-19T05:55:08Z

I was previously using the cxx11 ABI version of the libtorch package available on pytorch.org.
Just for an experiment, I used the Pre-cxx11 ABI version and now I am getting the following error:
Traceback (most recent call last):
  File "../../../test.py", line 14, in <module>
    main()
  File "../../../test.py", line 10, in main
    print(j.test(t1, t2))
TypeError: test(): incompatible function arguments. The following argument types are supported:
    1. (arg0: at::Tensor, arg1: at::Tensor) -> at::Tensor

Invoked with: tensor([[0.4624, 0.3462],
        [0.4250, 0.3891]]), tensor([[0.6797, 0.2709],
        [0.5302,  0.7420]])
Is this error related to #20356?
@anubane I'm getting the exact same error like this, "incompatible function arguments". Did you solve this by compiling the code via setup.py with CppExtension and BuildExtension in torch.utils.cpp_extension?

Yes, currently cpp_extension (setuptools) seems to be the only solution.

grybouilli · 2023-06-21T11:43:54Z

I was previously using the cxx11 ABI version of the libtorch package available on pytorch.org.
Just for an experiment, I used the Pre-cxx11 ABI version and now I am getting the following error:
Traceback (most recent call last):
  File "../../../test.py", line 14, in <module>
    main()
  File "../../../test.py", line 10, in main
    print(j.test(t1, t2))
TypeError: test(): incompatible function arguments. The following argument types are supported:
    1. (arg0: at::Tensor, arg1: at::Tensor) -> at::Tensor

Invoked with: tensor([[0.4624, 0.3462],
        [0.4250, 0.3891]]), tensor([[0.6797, 0.2709],
        [0.5302,  0.7420]])
Is this error related to #20356?
@anubane I'm getting the exact same error like this, "incompatible function arguments". Did you solve this by compiling the code via setup.py with CppExtension and BuildExtension in torch.utils.cpp_extension?
Yes, currently cpp_extension (setuptools) seems to be the only solution.

Hello, is there still no other solution than this ?

woywoy123 · 2023-06-21T19:33:07Z

Having the absolute same issue, but only for CUDA extensions.

grybouilli · 2023-06-21T20:27:47Z

Worked through my problem if this can help anyone:

Check that anything you're trying to build with libtorch is built with the same c++ standard lib that libtorch was built with.

Let me explain:
I was trying to integrate libtorch to a project that was compiled and built with clang - which uses libc++. Problem is, libtorch - unless built from source by your own hands - is built with gcc which uses libstdc++, so if you ever do what I did, you may as well try libtorch with or without abi, with or without _GLIBCXX_USE_CXX set to 0, you will get undefined symbols because the std::whatever::basic_string coming from your project get mixed up with the std::cxx11::string from libtorch.

So either build libtorch from source with your prefered compiler or build your project with the libstdc++ (gcc with c++11 standard and higher).

woywoy123 · 2023-06-21T20:52:28Z

https://github.com/r-barnes/pytorch_cmake_example

This looks very interesting.

alfie-nsugh · 2023-11-01T15:15:29Z

https://github.com/r-barnes/pytorch_cmake_example

This looks very interesting.

It is, thank you.

leeeizhang · 2025-05-16T10:10:18Z

Worked through my problem if this can help anyone:

Check that anything you're trying to build with libtorch is built with the same c++ standard lib that libtorch was built with.

Let me explain: I was trying to integrate libtorch to a project that was compiled and built with clang - which uses libc++. Problem is, libtorch - unless built from source by your own hands - is built with gcc which uses libstdc++, so if you ever do what I did, you may as well try libtorch with or without abi, with or without _GLIBCXX_USE_CXX set to 0, you will get undefined symbols because the std::whatever::basic_string coming from your project get mixed up with the std::cxx11::string from libtorch.

So either build libtorch from source with your prefered compiler or build your project with the libstdc++ (gcc with c++11 standard and higher).

thanks! It was very helpful. In the conda env, I can build cpp extension successfully only when libstdcxx-ng version is same with gcc.

conda create -n dev-test -c conda-forge gcc=11.1 gxx=11.1 libstdcxx-ng=11.1 python=3.11

gchanan added module: mkldnn Related to Intel IDEEP or oneDNN (a.k.a. mkldnn) integration module: cpp Related to C++ API triaged This issue has been looked at a team member, and triaged and prioritized into an appropriate module labels Feb 17, 2022

roastduck mentioned this issue Jun 23, 2022

PyTorch support cannot be built in Debug mode roastduck/FreeTensor#116

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Unable to build and use libtorch function via pybind11: undefined symbol error upon import #73016

Unable to build and use libtorch function via pybind11: undefined symbol error upon import #73016

Unable to build and use libtorch function via pybind11: undefined symbol error upon import #73016

Unable to build and use libtorch function via pybind11: undefined symbol error upon import #73016

Comments

🐛 Describe the bug

Versions