This repository provides the official PyTorch implementation of Stratified Transformer, a point-based approach to 3D point cloud semantic segmentation that uses standard multi-head self-attention to achieve a large receptive field and robust generalization. The method is notable as the first point-based approach to outperform voxel-based methods like SparseConvNet and MinkowskiNet. It includes memory-efficient CUDA kernels to handle variable-length tokens with shared memory acceleration.