[ICLR2026] LaplacianFormer: Rethinking Linear Attention with Laplacian Kernel

Implementation of "LaplacianFormer: Rethinking Linear Attention with Laplacian Kernel" (https://arxiv.org/pdf/2604.20368).

This repository provides the CUDA source code of the custom PyTorch operators introduced in the paper, including the Laplacian Subtraction kernel and the Newton Inverse kernel. These operators are the core building blocks of the Laplacian-kernel-based linear attention proposed in the paper.

Benchmark

The figure above shows the optimization effect of the provided CUDA operators.

🙈Note: The kernels in this repository are specifically tuned for the NVIDIA A100 (sm_80) GPU.

Citation

@inproceedings{zhe2026rethinking,
      title={LaplacianFormer: Rethinking Linear Attention with Laplacian Kernel},
      author={Zhe Feng and Sen Lian and Changwei Wang and Muyang Zhang and Tianlong Tan and Rongtao Xu and Weiliang Meng and Xiaopeng Zhang},
      year={2026},
      booktitle={ICLR},
}