faiss-wheels

faiss python wheel packages.

faiss

Overview

This repository provides scripts to build wheel packages for the faiss library.

Builds CPU-only version with cibuildwheel.
Bundles OpenBLAS in Linux/Windows
Uses Accelerate framework in macOS

There is also a source package to customize the build process.

Note GPU binary package is discontinued as of 1.7.3 release. Build a source package to support GPU features.

Install

Install the CPU-only binary package by:

pip install faiss-cpu

Note that the package name is faiss-cpu.

Supporting GPU or customized build configuration

The PyPI binary package does not support GPU. To support GPU methods or use faiss with different build configuration, build a source package. For building the source package, swig 3.0.12 or later needs to be available. Also, there should be all the required prerequisites for building faiss itself, such as nvcc and CUDA toolkit.

Building faiss

The source package assumes faiss is already built and installed in the system. If not done so elsewhere, build and install the faiss library first. The following example builds and installs faiss with GPU support and avx512 instruction set.

git clone https://github.com/facebookresearch/faiss.git
cd faiss
cmake . -B build -DFAISS_ENABLE_GPU=ON -DFAISS_ENABLE_PYTHON=OFF -DFAISS_OPT_LEVEL=avx512
cmake --build build --config Release -j
cmake --install build install
cd ..

See the official faiss installation instruction for more on how to build and install faiss.

Building a source package

Once faiss is built and installed, build the source package. The following builds and installs the faiss-cpu source package with GPU and AVX512.

export FAISS_ENABLE_GPU=ON FAISS_OPT_LEVEL=avx512
pip install --no-binary :all: faiss-cpu

There are a few environment variables that specifies build-time options.

FAISS_INSTALL_PREFIX: Specifies the install location of faiss library, default to /usr/local.
FAISS_OPT_LEVEL: Faiss SIMD optimization, one of generic, avx2, avx512. Note that AVX option is only available in x86_64 arch.
FAISS_ENABLE_GPU: Setting this variable to ON builds GPU wrappers. Set this variable if faiss is built with GPU support.
CUDA_HOME: Specifies CUDA install location for building GPU wrappers, default to /usr/local/cuda.

Development

This repository is intended to support PyPI distribution for the official faiss library. The repository contains the CI workflow based on cibuildwheel. Feel free to make a pull request to fix packaging problems.

Other relevant resources:

Packaging projects with GPU code

kyamagu / faiss-wheels

readme