license: cc-by-nc-nd-4.0

Introduction

TQCompressedGPT-2 is an advanced neural network model, offering a novel method for model compression through improved tensor decompositions. It addresses the challenges of computational and storage demands in NLP tasks, introducing a permutation-based enhancement to Kronecker decomposition, significantly reducing model size while maintaining performance.
TQCompressedGPT2 © 2024 by Terra Quantum AG is licensed under CC BY-NC-ND 4.0. To view a copy of this license, visit http://creativecommons.org/licenses/by-nc-nd/4.0/
Any entity who wishes to use this library for commercial purposes should contact info@terraquantum.swiss for more information.

Features

Model Size Reduction: Compresses the GPT-2small model from 124 million to 81 million parameters.
Permutation-Based Enhancement: Introduces a new permutation algorithm for matrix factorization, minimizing performance degradation.
Efficient Training Strategy: Employs multi-step knowledge distillation with a fraction (3.1%) of the OpenWebText dataset.
Performance: Outperforms DistilGPT-2 in comparative evaluations.

Permutation-Based Enhancement

In our work we employ permutation-based algorithm, which allows to achieve better decomposition approximation for weight matrices:

Methodology

For more details about the techniques of TQCompressedGPT-2, refer to our paper: (ADD LINK)TQCompressor: Improving Tensor Decomposition in Neural Networks via Permutations
TQCompressed Decomposition: Focuses on optimal permutation of weight matrices followed by Kronecker decomposition.
Knowledge Distillation: Uses an iterative compression method coupled with knowledge distillation, enhancing performance.
Application: Demonstrated on the GPT-2 model, showing its versatility and applicability to various neural network architectures.

Usage

The model and code are publicly available at:

Citation

If you find TQCompressedGPT-2 useful in your research, please cite the following paper:

@article{tqcompressedgpt2,
  title={TQCompressor: Improving Tensor Decomposition in Neural Networks via Permutations},
  author={Abronin, V., Naumov, A., Mazur, D., Bystrov, D., Tsarova, K., Melnikov, Ar., Oseledets, I., Dolgov, S., Brasher, R., Perelshtein, M.},
  journal={arXiv preprint arXiv:[insert_arxiv_id]},
  year={2023}
}

Acknowledgments

Terra Quantum AG, Kornhausstrasse 25, 9000 St. Gallen, Switzerland
Project contributors and researchers.