GElib is a C++/CUDA library for building equivariant neural networks. The library can be used as a pure C++ library or as a C++ extension for PyTorch. GElib uses cnine as its backend. Documentation for the C++ API can be found in the 'doc' directory. Documentation for the PyTorch API is at https://risi-kondor.github.io/GElib/.

GElib is released under a combination of the Mozilla Public License v. 2.0. and a custom noncommercial license to be found in the file NONCOMMERCIAL.TXT. The latter also applies to the cnine component. Commercial use requires a licence from the author(s).

Default installation as a PyTorch C++ extension

GElib is distributed in source code format, therefore to install it as a C++ extension you must have a working C++ compiler supporting C++17 (or higher) on your system (GCC or clang recommended). You also need PyTorch. If these requirements are satified, in most cases GElib can be installed simply by running

   pip install gelib

or

   pip3 install gelib

Installation with CUDA support

Compiling GElib with CUDA enabled is more complicated because it also requires the CUDA development environment to be present on your system. Further, you must use the same CUDA version as the version that PyTorch was compiled against. If, for example, your PyTorch needs CUDA 12.6, which is installed at /usr/local/cuda-12.6, you need to point the installation script to it with

   export CUDA_HOME="/usr/local/cuda-12.6"

To compile with CUDA enabled you must also set

   export WITH_CUDA="TRUE"

before running pip install gelib command.

Fallback installation method as a C++ extension

The fallback installation method is to clone the library with

   git clone --recurse-submodules git@github.com:risi-kondor/GElib.git

and install it manually by running

   pip install -e . --no-build-isolation

in its root directory. The --recurse-submodules flag ensures that a copy of cnine is bundled inside the GElib directory structure. The --no-build-isolation flag forces GElib to be built using versions of PyTorch etc. already present in your environment rather than temporary versions downloaded during the build process. This can help avoid version conflicts. The CUDA component of the library is enabled the same way as in the default installation method.

Credits

Lead developer: Risi Kondor

Contributors: Erik Henning Thiede, Ryan Keane