Awesome

<h1 align="center"> <p>🤗 GIFT-SW with PEFT</p></h1> <h3 align="center"> <p>GIFT-SW with State-of-the-art Parameter-Efficient Fine-Tuning (PEFT) methods</p> </h3>

This repository contains code for GIFT-SW method implemented with PEFT library. It could be used in the same interface as usual PEFT methods and easily pluggable into any code.

PEFT is integrated with Transformers for easy model training and inference, Diffusers for conveniently managing different adapters, and Accelerate for distributed training and inference for really big models.

Quickstart

Install PEFT directly from repository:

cd GIFT_SW/
pip install -e .

In case you have already installed PEFT, you will need to reinstall it:

cd GIFT_SW/
pip uninstall -y peft
pip install -e .

Get the activation scales for outlier computation from precomputed scales in QUIK repository or by collecting them with script from SmoothQuant

Prepare a model for training with a GIFT-SW method by wrapping the base model and PEFT configuration with get_peft_model.

from transformers import AutoModelForCausalLM
from peft import get_peft_config, get_peft_model, GIFTConfig, TaskType

model_name_or_path = "facebook/opt-1.3b"
tokenizer_name_or_path = "facebook/opt-1.3b"
path_to_act_scales = "./opt-1.3b.pt"

peft_config = GIFTConfig(
    outlier_num=64,
    target_modules=['q_proj', 'k_proj'],
    path_to_act_scales=path_to_act_scales,
)

model = AutoModelForCausalLM.from_pretrained(model_name_or_path)
model = get_peft_model(model, peft_config)
model.print_trainable_parameters()
#"trainable params: 6,291,456 || all params: 1,517,182,976 || trainable%: 0.4147"

To save and later inference GIFT-SW model it is highly recommended to "merge_and_unload()" the model as GIFT-SW is not a regular adapter, but a learned subset of model weights. Further tuning of already trained GIFT-SW model is equivalent to merging the model and learning new one.

How to get activations in SmoothQuant

To get the activation scales for your model you will need to get them with SmoothQuant method, it is simple and easy to use.

git clone https://github.com/mit-han-lab/smoothquant

# make sure the git-lfs is installed
# curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh | bash
# apt-get install git-lfs
# git lfs install

# clone the calibration data
git clone https://huggingface.co/datasets/mit-han-lab/pile-val-backup
# move to smoothquant and run the script

cd smoothquant
python examples/generate_act_scales.py \
    --model-name <your model name> \
    --output-path <save path .pt> \
    --num-samples 512 \ #number of calibration samples
    --seq-len 2048 \ #max sequence length
    --dataset-path ../pile-val-backup/val.jsonl.zst

Citing GIFT-SW

To use GIFT-SW in your publication, please cite it by using the following BibTeX entry.

@misc{zhelnin2024giftswgaussiannoiseinjected,
      title={GIFT-SW: Gaussian noise Injected Fine-Tuning of Salient Weights for LLMs}, 
      author={Maxim Zhelnin and Viktor Moskvoretskii and Egor Shvetsov and Egor Venediktov and Mariya Krylova and Aleksandr Zuev and Evgeny Burnaev},
      year={2024},
      eprint={2408.15300},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2408.15300}, 
}