Are the PyPi wheels built with CUDA?

[pbsds]

This section of the readme is not clear about this

pycolmap/README.md

Lines 69 to 72 in 05abbd0

	PyCOLMAP can leverage the GPU for feature extraction, matching, and multi-view stereo if COLMAP was compiled with CUDA support.
	Similarly, PyCOLMAP can run Delauney Triangulation if COLMAP was compiled with CGAL support.
	This requires to build the package from source and is not available with the PyPI wheels.

We know it is not built with CGAL, but CUDA is not clearly said if it supported on PyPI, and if it is, which version? CUDA 10? CUDA 11?

[sarlinpe]

PyPI wheels are not built with CUDA support at the moment. You are very welcome to help us support it. Some pointers:

• We'd need to install the Cuda toolkit in the CenOS manylinux image, either through yum (here is a good starting point) or via pip (described here)
• It would be nice to not package the Cuda libs in the wheel and instead depend on Cuda pip packages - we'd need to add these requirements to the pycolmap wheels.
• I don't know whether we need to build a separate wheel for each version of Cuda

[pbsds]

I think the pytorch way, where a MAJOR.MINOR.PATCH+LOCAL scheme where LOCAL is e.g. cuda11 probably is the way to go. I'm not asking for you to dist cuda wheel, but I am hoping for you to clarify the state on PyPi in the README. 😄

[sarlinpe]

Good point, fixed by #188.

[Phil26AT]

Just a note: The pycolmap conda wheels are built with CUDA.

[sarlinpe]

Indeed, the build repo is here: https://github.com/conda-forge/pycolmap-feedstock I am not familiar with the conda build CI