I am sure the -D_GLIBCXX_USE_CXX11_ABI=0 has been discussed enough times and we have to support it for the early manylinux PEP513 compatible installs.
However when digging through the builder scripts I noticed we have a new toolchain since last year built with no_cxx11_abi for the compiler itself (Comparing pytorch:main...bertmaher:llvm_cxx11_abi · pytorch/builder · GitHub)
Here is an example of using a “regular” cxx11 compiler (still provided by the pytorch-manylinux-cpu docker image) linking against a released pre-cxx11 abi Pytorch wheel. [release] fix cuda versions · pytorch/functorch@7f23d3d · GitHub
We are trying to find the right integration points for torch-mlir and wanted to understand the need to build the compiler itself with -D_GLIBCXX_USE_CXX11_ABI=0 vs just building Pytorch with those flags and correct PYBIND11 Internal flags.
@bertmaher @malfet Any thoughts ? Thanks thanks