This repository contains the artifact for our ASPLOS 2026 paper: "RedFuser: An Automatic Operator Fusion Framework for Cascaded Reductions on AI Accelerators".

RedFuser is a novel framework for optimizing cascaded reductions in deep learning compilers. Built on top of Apache TVM, RedFuser introduces a series of compiler transformation passes that enable efficient fusion of reduction operations with other computations, particularly targeting modern GPU architectures.

• [2026-03]: Support Flash Decoding, Moe Routing and Quant GEMM examples.
• [2026-01]: RedFuser is now avaliable with flash-attention example.
• [2025-11]: 🎉 RedFuser is accepted by ASPLOS 2026!

• flash-attention
• flash-decoding
• moe-routing
• fp8 quant+gemm

Please follow https://tvm.apache.org/docs/install/index.html to install.

For flash-attention example, see python/tvm/redfuser/example/flash_attention.py.

redfuser/
├── python/tvm/redfuser/       # RedFuser Python implementation
│   ├── transform/             # Core transformation passes
│   └── example/               # Example workloads
│ ...

If you use RedFuser in your research, please cite our paper:

@article{RedFuser,
  title={RedFuser: An Automatic Operator Fusion Framework for Cascaded Reductions on AI Accelerators},
  author={Xinsheng Tang and Yangchen Li and Nan Wang and Zhiyi Shu and Xingyu Ling and Junna Xing and Peng Zhou and Qiang Liu},
  journal={Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2},
  year={2026},
  url={https://arxiv.org/abs/2603.10026}
}

RedFuser is licensed under the Apache License 2.0.

This project builds upon Apache TVM. We thank the TVM community for their excellent infrastructure and support.