https://github.com/gunale0926/sorsa

SORSA: Singular Values and Orthonormal Regularized Singular Vectors Adaptation of Large Language Models
https://github.com/gunale0926/sorsa

deep-learning fine-tuning llama lora machine-learning nlp peft python pytorch rwkv sorsa svd transformer

Last synced: 19 days ago
JSON representation

SORSA: Singular Values and Orthonormal Regularized Singular Vectors Adaptation of Large Language Models

• Host: GitHub
• URL: https://github.com/gunale0926/sorsa
• Owner: Gunale0926
• License: apache-2.0
• Created: 2024-06-09T19:28:18.000Z (5 months ago)
• Default Branch: main
• Last Pushed: 2024-10-16T01:17:23.000Z (about 1 month ago)
• Last Synced: 2024-10-22T23:28:03.784Z (23 days ago)
• Topics: deep-learning, fine-tuning, llama, lora, machine-learning, nlp, peft, python, pytorch, rwkv, sorsa, svd, transformer
• Language: Python
• Homepage: https://arxiv.org/abs/2409.00055
• Size: 4.38 MB
• Stars: 43
• Watchers: 2
• Forks: 2
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE
  • Citation: CITATION.cff

Awesome Lists containing this project

README

        
# SORSA: Singular Values and Orthonormal Regularized Singular Vectors Adaptation of Large Language Models

[![arXiv](https://img.shields.io/badge/arXiv-2409.00055-b31b1b.svg)](https://arxiv.org/abs/2409.00055) [![PyPI - Version](https://img.shields.io/pypi/v/sorsa)](https://pypi.org/project/sorsa/)

This repository contains the codes of experiments of the paper *SORSA: Singular Values and Orthonormal Regularized Singular Vectors Adaptation of Large Language Models*.

![Architecture of SORSA](./assets/SORSA.png)

The rapid advancement in large language models (LLMs) comes with a significant increase in their parameter size, presenting challenges for adaptation and fine-tuning. Parameter-efficient fine-tuning (PEFT) methods are widely used to adapt LLMs for downstream tasks efficiently. In this paper, we propose Singular Values and Orthonormal Regularized Singular Vectors Adaptation, or SORSA, a novel PEFT method. Each SORSA adapter consists of two main parts: trainable principal singular weights $W_p = U_p \text{diag}(S_p) V^\top_p$, and frozen residual weights $W_r = U_r \text{diag}(S_r) V^\top_r$. These parts are initialized by performing SVD on pre-trained weights. Moreover, we implement and analyze an orthonormal regularizer. SORSA adapters could be merged during inference, thus eliminating any inference latency.

## Empirical Experiments

![Exp](./assets/Exp.png)

## Reproduce the Experiments

First, install `sorsa` package from pip:

```bash

pip install sorsa

```

Then, create `.env` file in the root directory of the project and add your [Hugging Face Access Token](https://huggingface.co/settings/tokens):

```bash

hf=Your_Hugging_Face_Access_Token

```

### Llama 2 7B, Mistral v0.1 7B and Gemma 7B

First, install the packages via anaconda

```bash

conda env create -f environment.yml

```

Run scripts from `./scripts/train_sorsa.sh` to train the model.

After training, run the ``./scripts/merge_sorsa.sh`` to merge the adapter to the base model:

Run following command to evaluate on GSM-8K:

```bash

python3 run.py --name llama2_sorsa_r128 \

  --test \

  --test-dataset gsm-8k \

  --test-precision bf16

```

Run following command to evaluate on MATH:

```bash

python3 run.py --name llama2_sorsa_r128 \

  --test \

  --test-dataset math \

  --test-precision bf16

```

Run following command to evaluate on HumanEval:

```bash

python3 run.py --name llama2_sorsa_r128 \

  --test \

  --test-dataset humaneval \

  --test-precision bf16

```

### RWKV6

If you are training, merging or testing RWKV6 model, please add `--rwkv` flag to `run.py`.

## Cite the work

You could cite the work by using the BibTeX code as follows:

```bibtex

@misc{CaoSORSA2024,

  title = {{{SORSA}}: {{Singular Values}} and {{Orthonormal Regularized Singular Vectors Adaptation}} of {{Large Language Models}}},

  shorttitle = {{{SORSA}}},

  author = {Cao, Yang},

  year = {2024},

  month = oct,

  number = {arXiv:2409.00055},

  eprint = {2409.00055},

  primaryclass = {cs},

  doi = {10.48550/arXiv.2409.00055},

  url = {http://arxiv.org/abs/2409.00055},

  archiveprefix = {arXiv},

  keywords = {Computer Science - Computation and Language,Computer Science - Machine Learning},

}

```