nerfstudio-project
document installation from published source package to compile CUDA during the installation
In the documentation it is said one should use pip install git+https://github.com/nerfstudio-project/gsplat.git
to build the CUDA code during installation.
An alternative that allows one to build the CUDA code during installation but use the published package from pypi.org is to use pip install --no-binary=gsplat gsplat which will do the installation using the source package instead of the wheel file
it would be great to document this option too.
related to this point, I would be curious to understand the motivation behind publishing wheels without pre-build binaries on pypi.org. My understanding is that in general wheels are expected to contain binaries. Having wheels without the pre-compiled binaries seems a bit confusing and somehow makes the installation process more complex on windows because one need then to make sure cl.exe is on the path at runtime.
the possibility to install with pip install --no-binary=gsplat gsplat has been added to the windows installation documentation in this PR https://github.com/nerfstudio-project/gsplat/pull/365. It could also be added on the main readme page too