Neighborhood Attention Transformer (NAT) and its dilated variant (DNAT) are vision transformer architectures that replace global self-attention with localized neighborhood attention for improved efficiency. The models achieve state-of-the-art performance on multiple computer vision benchmarks including instance segmentation, semantic segmentation, and panoptic segmentation on ADE20K, Cityscapes, and COCO datasets. The implementation includes PyTorch models and a custom CUDA extension (NATTEN) for accelerated neighborhood attention computation.