ImportError raised when testing wheels on windows and macos

[robomics]

Thanks for the very useful tool!

I am struggling to debug an ImportError that is raised when building and testing a minimal pybind11 project using cibuildwheel.

CI config: link
Build log: link

The import error reads something like: ModuleNotFoundError: No module named 'mymodule.mymodule' (I am not sure where the second mymodule is coming from).

The project itself is setup as suggested by https://github.com/pybind/cmake_example.

The only differences being:

• setup.py is calling conan to install pybind11 before building the project
• setup.py creates the following __init__.py file:

from .mymodule import call_me

try:
    from importlib.metadata import version
except ModuleNotFoundError:
    from importlib_metadata import version

__version__ = version("mymodule")

Oddly enough the tests pass on Ubuntu (see here).
While debugging the error, I've setup a simple GHA workflow that installs the package using pip, then runs the tests with pytest on Ubuntu, macOS and Windows.
The workflow can be found here.
To my surprise there the tests pass on all platforms without raising ImportErrors (see here).

At this point I am no longer sure whether the issue lies with cibuildwheel or with the way I am setting up the project.

Any help is appreciated!

[Czaki]

The import error reads something like: ModuleNotFoundError: No module named 'mymodule.mymodule' (I am not sure where the second mymodule is coming from).

as package is named mymodule them from .mymodule import call_me evaluates to from mymodule.mymodule import call_me

On Ubuntu build, contains binary file named mymodule:

You may try to disable tests and download created wheel to manually inspect them.

[robomics]

Thanks for your rapid reply!
I've followed your advice and temporarily disabled tests.
The wheels look ok to me (they all have proper names and contain an __init__.py file and a shared library).
I have successfully installed one of the Windows wheels in a VM, and was able to import the library and call mymodule.call_me() without errors, so I am still unsure why the wheels lead to import errors when installed by cibuildwheel.

[henryiii]

I am not sure where the second mymodule is coming from

You have the second one in your __init__.py. .mymodule means mymodule from the current module, which is mymodule/. So that means you expect mymodule/mymodule.<plat>.so.

Does your shared library sit inside a matching folder, next to the __init__.py?

I would highly recommend pybind/scikit_build_example be followed instead of pybind/cmake_example, which is very fragile, by the way.

[henryiii]

# Setuptools bug causes collision between pypy and cpython artifacts

That's not true anymore, it was fixed, by the way.

[henryiii]

IMO it looks like it's not being installed into the proper folder. It's much better to do it the scikit-build-core way, with the 'install' CMake command. CMake fixes up the paths when you use that, and you can tell it what folder to install to. You are currently trying to ship the build folder, which is not recommended.

[robomics]

Thanks for your suggestions!
I've updated the project to use scikit-build-core and things are now running smoothly.