cnine

Lightweight C++ tensor library

Cnine is a simple C++/CUDA tensor library developed by Risi Kondor's group at the University of Chicago. Cnine is designed to make some of the power of modern GPU architectures accessible directly from C++ code, without relying on complex proprietary libraries.

Documentation for the Python/PyTorch API is at https://risi-kondor.github.io/cnine/.

Cnine is released under the custom noncommercial license included in the file LICENSE.TXT

Installation

GPU / CUDA installation

Please install the CUDA toolkit first!

Install with pip (includes PyTorch 2.0+):

pip install .

CPU installation

For custom PyTorch versions:

1. Install desired PyTorch CPU first

   pip install torch --index-url https://download.pytorch.org/whl/cpu
2. Then install the build dependencies of cnine manually. (Listed in pyproject.toml/build-system/requires.) Except for torch since we manually installed that first.

   pip install scikit-build-core pybind11
3. Then install cnine, we disable the isolation build, so that module we just manually installed are used as dependencies during the build stage

   pip install --no-build-isolation .

Manual CMake Build

Cnine uses scikit-build-core. To build manually:

git clone https://github.com/risi-kondor/cnine.git
cd cnine
mkdir build && cd build
cmake .. -DCMAKE_PREFIX_PATH="/path/to/python/installation"
make -j4