/RetNet

An implementation of "Retentive Network: A Successor to Transformer for Large Language Models"

Primary LanguagePythonMIT LicenseMIT

RetNet

An implementation of Retentive Network: A Successor to Transformer for Large Language Models in PyTorch.

About this repository

This is a minimal, pure pytorch implementation of RetNet. RetNet paper: Retentive Network: A Successor to Transformer for Large Language Models.

The contributors(s) to this repository are not authors of the original paper. All credit for the idea and formulation of RetNet goes to the original authors.

The purpose of this repository is to aid scientific and technological understanding and advancement. The code prioritizes correctness and readability over optimization.

Features implemented

  • Single-scale and MultiScale retention:
    • parallel paradigm
    • recurrent paradigm
    • chunkwise paradigm
  • Multi-layer retentive network with FFN and LayerNorm
    • parallel paradigm
    • recurrent paradigm
    • chunkwise paradigm
  • Causal language model (CLM) built on top of the the retentive network

Usage and Examples:

  • See scripts prefixed with test_ for examples of basic usage

Positional Encodings

The main implementation in src/ uses Microsoft's xPos for positional encoding.

The implementation in src/complex uses complex values to encode position, which requires parameter and data throughput types to be torch.ComplexFloat (64-bit). This has some limitations due to there not yet being torch support for half-precision complex types. It also requires twice the amount of memory as real-valued data at 32-bit precision.

Contributions

All contributions are welcome. Please see issues for an idea of what needs doing.

If you would like to contribute to this project, please fork it and submit a pull request for review.

References

@misc{sun2023retentive,
      title={Retentive Network: A Successor to Transformer for Large Language Models}, 
      author={Yutao Sun and Li Dong and Shaohan Huang and Shuming Ma and Yuqing Xia and Jilong Xue and Jianyong Wang and Furu Wei},
      year={2023},
      eprint={2307.08621},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}