Issue with complex case of using PyTorch with PyBind

[WillemvJ]

I am attempting to use libtorch (the C++ pytorch front-end) in a pybind extension, but I cannot get it to consistently work accross platforms. Any help or ideas are appreciated.

My code (cmake) add a module like so

pybind11_add_module(DP_bindings)

and find and links libtorch from anaconda:

"C:/Users//Anaconda3/envs/envDP/Lib/site-packages/torch/share/cmake"

The bindings are just bind a function that represents a simple call to some libtorch function (for now).

On windows, I can successfully import the bindings as long as I use the "envDP" python version , i.e. the one I linked my libtorch extension from, and as long as I call:

import torch
import DP_bindings

(If I don't first import torch, I get "DLL load failed while importing DP_Bindings: The specified module could not be found." )

On Linux, however, the same calls yield errors with undefined symbols:

/gpfs/home1/..../DP_Bindings.cpython-311-x86_64-linux-gnu.so: undefined symbol: _ZN8DynaPlex6ParamsC1ESt16initializer_listISt5tupleIJSsSt7variantIJbDnldSsS0_St6vectorIlSaIlEES4_IdSaIdEES4_ISsSaISsEES4_IS0_SaIS0_EEEEEEE

These errors can be avoided by ensuring that the cmake defines only a single target (DP_bindings). However, when adding a second target (say an OBJECT or a static or dynamic library), the errors consistently appear. I suspect issues with ABI compatibility. But it could be something else.

Related:
#2113
(I studied their setup. In my case, I could succesfully use multiple targets, but not if I also use external libraries like libtorch.)

Also, maybe this is related: #3793

I am happy to try a downsized example with code, if somebody with deep pybind knowledge is interested to investigate. I would also be happy with any suggestion on other things to try.