(This will become) An reimplementation of Retentive Network: A Successor to Transformer for Large Language Models in Keras Core/Keras 3.0 to make it backend-agnostic and provide the functionality to serialize and export models making use of xPos encoding and the custom Retention layers, forked from Jamie Stirling's native PyTorch implementation: https://github.com/Jamie-Stirling/RetNet.
This is Keras Core/Keras 3.0 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. The purpose of this fork is to provide the ability to build RetNet-like models in Keras Core/Keras 3.0, which will allow free choice of backend to use (PyTorch and TensorFlow are the main targets, JAX is for now a maybe).
- 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
- See scripts prefixed with
test_
for examples of basic usage
The main implementation in src/
uses Microsoft's xPos (refactored for Keras Core/Keras 3.0) 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.
All contributions are welcome.
If you would like to contribute to this project, please fork it and submit a pull request for review.
@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}
}