cuequivariance-torch

You can run equivariant tensor-product computations on GPUs by importing cuequivariance_torch and instantiating SegmentedPolynomial-based modules with CUDA acceleration, integrating them directly into your PyTorch model workflows.

Available equivariant neural network components include ChannelWiseTensorProduct, FullyConnectedTensorProduct, Linear, SphericalHarmonics, Rotation, Inversion, and SymmetricContraction, plus layers like BatchNorm and FullyConnectedTensorProductConv.

SegmentedPolynomial backends include naive, uniform_1d, fused_tp, and indexed_linear, with CUDA acceleration applied automatically when available to optimize equivariant operations performance.

Yes, you need both cuequivariance and cuequivariance_torch installed to execute equivariant tensor-product computations and build neural network components within your PyTorch environment.

Yes, components like FullyConnectedTensorProductConv and BatchNorm are ready-to-use layers designed for model building and training, plugging directly into standard PyTorch workflows with GPU acceleration.

The fused_tp backend is typically preferred for performance, utilizing CUDA acceleration for equivariant tensor products, while the naive backend serves as a reference or fallback when specialized GPU kernels are unavailable.