Welcome to the TyphoonMLA Community Repository!

For more details on TyphoonMLA method, check out our preprint paper.

Run TyphoonMLA on your own machine or on a server.

Results are collected from community contributors. Feel free to create a pull request with your results from different platforms. We will show all benchmark results in the main page.

How to run on your local machine

Start a docker container from our pre-built image:

docker run -it --rm --gpus=all --runtime=nvidia --name tree-mla acyuzuguler/tree-mla:latest /bin/bash

Run experiments and plot results:

bash run.sh

How to run it on public cloud infra

Start a VM with the docker image acyuzuguler/tree-mla:latest

Run experiments and plot results:

bash run.sh

To rebuilt image from scratch

docker build -f dockerfiles/Dockerfile -t tree-mla:latest .

Results from community contributors

We welcome results & feedback from the community. If you run this code in a new architecture, please add your results below here and create a PR.

How to cite this work

@misc{yuzuguler2025typhoonmla,
      title={TyphoonMLA: A Mixed Naive-Absorb MLA Kernel For Shared Prefix}, 
      author={Ahmet Caner Yüzügüler and Ahmet Çelik and Jiawei Zhuang and Lukas Cavigelli},
      year={2025},
      eprint={2509.21081},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2509.21081}, 
}