Attention Gym

September 7, 2026 · View on GitHub

Attention Gym is a collection of kernels, guides, and examples for FlexAttention and other novel attention variants.

📚 Docs | 🎯 Features | 🚀 Getting Started | 💻 Usage | 🛠️ Dev | 🤝 Contributing | ⚖️ License

📖 Overview

Attention Gym began as a library of examples showing the many ways to express attention variants with the FlexAttention API. It is growing into a broader playground for attention, with the addition of sparse attention kernels, linear attention APIs for training and inference as well as showcasing how to use FlexAttention and friends in real workloads.

Two new Flexys—Sparse Attention and Linear Attention—enter Attention Gym while the original Flexys train

🎯 Features

  • FlexAttention masks and score modifications
  • Sparse attention patterns
  • APIs and kernels for efficient GDN and KDA
  • Utility functions for creating and combining attention masks
  • Examples of how to use FlexAttention in real-world scenarios

🚀 Getting Started

Prerequisites

  • PyTorch (version 2.5 or higher)

Installation

Install the official wheel from PyPI:

pip install attn-gym

The base package intentionally keeps its runtime dependency surface small: it depends only on PyTorch (unpinned). Optional features live behind extras, so install only what you need:

pip install "attn-gym[linear]"  # Linear-attention APIs and kernels
pip install "attn-gym[viz]"     # Visualization and example dependencies

Warning

Attention Gym is under active development. We reserve the right to make backward-incompatible changes between releases. If you depend on a particular API or kernel behavior, hard-pin the version you test, for example: pip install "attn-gym[linear]==X.Y.Z".

💻 Usage

Attention Gym supports three complementary workflows:

  1. Compose FlexAttention building blocks. Import mask_mod and score_mod functions and pass them directly to PyTorch's FlexAttention APIs.
  2. Use sparse and linear-attention APIs and kernels. Build with selected_attention, GDN and KDA chunk, recurrent, and decode paths, and short-convolution primitives. See the compressed sparse attention and delta-rule (KDA/GDN) training for working examples.
  3. Run real workloads and benchmarks. The examples/ directory covers paged, ring, and variable sparse attention, CUDA Graphs, determinism, compilation, and profiling. Most of this should serve as inspiration for fun things you might build from our building blocks :)

🛠️ Dev

Install dev requirements

pip install -e ".[dev]"

Install and run the repository hooks:

prek install
prek run --all-files

🤝 Contributing

We welcome contributions to Attention Gym, especially new Masks or score mods! Here's how you can contribute:

Contributing Mods

  1. Create a new file in the attn_gym/masks/ for mask_mods or attn_gym/mods/ for score_mods.
  2. Implement your function, and add a simple main function that showcases your new function.
  3. Update the attn_gym/*/__init__.py file to include your new function.
  4. Optionally, add an end-to-end example using your new function in the examples/ directory.

See CONTRIBUTING.md for more details.

⚖️ License

Attention Gym-authored code is released under the BSD 3-Clause License. Vendored third-party components retain the licenses and notices included alongside their source.