Dynamic linking

If you want to support dynamic linkages between python projects or system libraries, you will likely encounter some issues in making sure the compiled libraries/python bindings work after the wheel is created and the python project is installed on the system. The most common issue are the missing hints pointing to where the runtime libraries are located, specifically RPATH on Linux and MacOS systems, and PATH/os.add_dll_directory on Windows systems. Here are some recommendations on how to address them.

Wheel repair tools

There are wheel repair tools for each operating system that bundle any dynamic libraries used and patch the libraries/python bindings to point to prioritize those libraries. The most common tools for these are auditwheel for Linux, delocate for MacOS and delvewheel for Windows. cibuildwheel incorporates these tools in its repair wheel feature.

These tools also rename the library with a unique hash to avoid any potential name collision if the same library is being bundled by a different package, and check if the packages confirm to standards like PEP600 (manylinux_X_Y). These tools do not allow to have cross wheel library dependency.

Manual patching

You can manually make a relative RPath. This has the benefit of working when not running scikit-build-core, as well.

The RPATH patching can be done as

if(APPLE)
  set(origin_token "@loader_path")
else()
  set(origin_token "$ORIGIN")
endif()
set_property(TARGET <target> PROPERTY INSTALL_RPATH
  "${origin_token}/install_path/to/dynamic_library"
)

For Windows patching, this has to be done at the python files using os.add_dll_directory at the top-most package __init__.py file or top-level python module files.

import os
from pathlib import Path

dependency_dll_path = Path(__file__).parent / "install_path/to/dynamic_library"
os.add_dll_directory(str(dependency_dll_path))