Instructions to raise PR for addition of shared library files(.so) and .cpp files

[vineel96]

Feature request

Hello @ArthurZucker @zucchini-nlp @Rocketknight1 @gante,
I have an proposal for improvement of Mamba model for which I wanted to raise PR.
Method: The code changes are made in c++ file. I use cython to integrate c++ code to python. For this I will be generating an shared library file (.so) which will have the code. I will import it as library in transformers python code(Ex: import library) and call API's (Ex: library.api_name() ) from it.

I have queries on how to raise a PR for the above scenario:

1. In which folder path should I keep the cython c++ binding codes ?
2. How to provide the share library file (.so) file in the PR, In which location should I keep?

Motivation

I will be detailing the motivation when I raise the PR for it. Now I need to instructions to raise the PR

Your contribution

I will be submitting a PR.

[vineel96]

@ArthurZucker @zucchini-nlp @Rocketknight1 @gante, Any help regarding this?

[zucchini-nlp]

We usually have kernels for CUDA/CPU here https://github.com/huggingface/transformers/tree/main/src/transformers/kernels, so I believe mamba will also belong there

[vineel96]

@zucchini-nlp , To bind c++ code with python this repo is using torch.utils.cpp_extension to load ".cpp" extension files and create shared file(.so) (https://github.com/huggingface/transformers/blob/main/src/transformers/models/mra/modeling_mra.py#L59 ) . But torch.utils.cpp_extension uses pybindings to bind c++ code with python. In our experiments, we found that cython is faster than pybindings when we call c++ code from python. So can we integrate using cython? or it is mandatory to follow what transformers repo uses by default?