NVIDIA/cuEquivariance

cuEquivariance is a math library that is a collective of low-level primitives and tensor ops to accelerate widely-used models, like DiffDock, MACE, Allegro and NEQUIP, based on equivariant neural networks. Also includes kernels for accelerated structure prediction.

What it solves

It simplifies the creation of high-performance geometric neural networks that respect symmetries (equivariance), which is essential for building robust physical models that are data-efficient in 3D space.

How it works

The library provides a comprehensive API for describing segmented polynomials (built from segmented tensor products) and uses optimized CUDA kernels to execute these operations efficiently. It offers native bindings for both PyTorch and JAX to allow seamless integration into machine learning workflows.

Who it’s for

Researchers and developers building geometric neural networks and physical models that require rotation and translation equivariance in three-dimensional space.

Highlights

  • High-performance execution via optimized CUDA kernels.
  • Support for segmented polynomials and triangular operations.
  • Compatible with both PyTorch and JAX.
  • Designed for data-efficient AI models that respect physical symmetries.

Related

  • Project
    facebookresearch/pytorch3d

    A PyTorch-based library for 3D Computer Vision research providing differentiable rendering and efficient tools for manipulating 3D meshes and point clouds.

  • Project
    NVIDIA/apex

    A collection of NVIDIA-maintained utilities for PyTorch that streamline mixed precision and distributed training through optimized CUDA extensions.

  • Project
    qubvel-org/segmentation_models.pytorch

    A PyTorch library providing a high-level API for image semantic segmentation, featuring 12 architectures and over 800 pretrained encoders.

  • Project
    facebookincubator/AITemplate

    A Python framework that transforms deep neural networks into optimized CUDA/HIP C++ code for high-performance inference serving on NVIDIA and AMD GPUs.

  • Project
    facebookresearch/xformers

    A toolbox of optimized, customizable building blocks for Transformer models that accelerates research by providing memory-efficient and high-performance CUDA kernels.