Why is LibTorch from pip not build with CXX11 ABI

The opencv and torch problem continues to persist with the libtorch that comes torch installed via pip. I use conan as my package manager for most packages, so naturally i install opencv via conan. But when i try to link both torch installed via pip (cmake … -DCMAKE_BUILD_TYPE=Release -DCMAKE_PREFIX_PATH=python3 -c 'import torch;print(torch.utils.cmake_prefix_path)' …) and opencv from conan i get the usual error that everyone faces. I need to specify in my conan profile the cxx flag CXXFLAGS=-D_GLIBCXX_USE_CXX11_ABI=0 which builds opencv without the cxx11 abi and then everything compiles as expected. I have seen many threads that suggest to use the libtorch cpp package from the site with abi libtorch-cxx11-abi-shared-with-deps (see Library conflict between libtorch and opencv? - #8 by lzl_jnu), but i like the idea of using the libtorch that comes with pip, it would reduce duplicates of the package. I was wondering why the libtorch that comes torch pip is not compile with -D_GLIBCXX_USE_CXX11_ABI=1 ? I use opencv/4.8.1 (conan) and torch==2.4.0 (pip). My conan profile for the curious:

[settings]
os=Linux
arch=x86_64
compiler=gcc
compiler.version=11
compiler.libcxx=libstdc++11
compiler.cppstd=17
build_type=Release

[options]

[tool_requires]

[buildenv]
CC=/usr/bin/gcc-11
CXX=/usr/bin/g++-11
CXXFLAGS=-D_GLIBCXX_USE_CXX11_ABI=0 # I need it for the moment to tell conan to install opencv withouth cxx abi

As you can see i am pretty up to date with gcc version and all. So just curious, is it a miss that the libtorch packaged with torch pip is not the one installed with abi, or is it because it would break something?

The wheels support manylinux based on CentOS7 as described here and thus are build without the CXX11 ABI. Since CentOS7 is EOL already, future releases could relax the requirement and use the CXX11 ABI.

Ah i see, seems logical. I guess i’ll wait. It seems to be coming soon enough. Thanks