pip install -e does not build shared libraries

[Cebtenzzre]

When I try to use `pip install -e .` with llama-cpp-python or
bitsandbytes, it doesn't build the shared libraries like `pip install
.` does.

In the case of llama-cpp-python, there is no _skbuild folder generated
at all, just llama_cpp_python.egg-info. This results in an error when
I try to import it, e.g. "FileNotFoundError: Shared library with base
name 'llama' not found". In this case, `python setup.py develop` also
works.

In the case of bitsandbytes, there is no build folder generated at
all, just bitsandbytes.egg-info. This results in an error when I try
to import it, e.g. "CUDA SETUP: Required library version not found:
libbitsandbytes_cuda121_nocublaslt.so. Maybe you need to compile it
from source?". In this case, `python setup.py develop` does not seem
to work either.

Is this something that can be fixed in the projects' setup.py or
pyproject.toml, or is it an issue with pip or setuptools?

Thanks,
Cebtenzzre

[Cebtenzzre]

On Wed, Jun 21, 2023 at 5:45 PM Cebtenzzre <cebte...@gmail.com> wrote:
> In the case of bitsandbytes, there is no build folder generated at
> all, just bitsandbytes.egg-info. This results in an error when I try
> to import it, e.g. "CUDA SETUP: Required library version not found:
> libbitsandbytes_cuda121_nocublaslt.so. Maybe you need to compile it
> from source?". In this case, `python setup.py develop` does not seem
> to work either.

Correction: `make` must be run manually to build bitsandbytes. I made
a mistake while testing. `make clean` doesn't work correctly, so I had
to use `git clean` to get repeatable results.

[Cebtenzzre]

It turns out that this is a scikit-build issue:
https://github.com/scikit-build/scikit-build/issues/740
https://github.com/scikit-build/scikit-build/issues/939
https://github.com/scikit-build/scikit-build/issues/981