Flash EQ-Linear: Accelerating Equivariant Linear Layers via Group-wise Discrete Fourier Transform

Researchers introduced Flash EQ-Linear, a CUDA-based method that speeds up equivariant linear layers with group-wise Fourier techniques while preserving exact outputs.

The layer rewrite turns equivariant operations into circular convolution plus channel transforms, enabling efficient acceleration with custom kernels and real DFT.

It delivers notable inference speedups in standalone layers and in full models such as Flash EQ-ViT and Flash EQ-Swin.

The work shows equivariant vision networks can improve compute efficiency, not just parameter efficiency, making them more practical and competitive.