nd-rotary-encodings¶

Welcome to the documentation for nd-rotary-encodings.

This repository grew out of an effort at Berkeley Lab towards applying modern computer vision techniques like Detection Transformers to large, sparse scientific images. Since scientific images like electron microscope iamges can be very large with very small objects, a key requirement was to be able to process high-resolution images without rescaling them to standard sizes (e.g., a few hundred pixels in width and height) typically used for computer vision.

The nd-rotary-encodings repository features highly-optimized PyTorch kernels and layers for RoPE-encoding N-dimensional sequences, with exceptional scaling improvements over standard implementations.