[EMNLP 2024 Findings] QEFT: Quantization for Efficient Fine-Tuning of LLMs

This is the code for the paper QEFT: Quantization for Efficient Fine-Tuning of LLMs.

Table of contents

• Install
• Usage

Install

We highly recommend using a Docker image that supports CUDA. If you prefer Anaconda, you need to set up CUDA for kernel use.

0. A) Using Docker

docker run -it --gpus all --ipc=host -v {local_storage}:{docker_container_storage} pytorch/pytorch:2.0.0-cuda11.7-cudnn8-devel

# install git
apt update && apt install git -y

0. B) Using Anaconda instead of Docker

conda create -n owq python=3.10 -y
conda activate owq

1. Clone the QEFT repository

git clone https://github.com/xvyaward/qeft
cd QEFT_PV

2. Install all the dependencies

pip install -e .

3. Install the OWQ CUDA kernel

cd qeft/kernel
python setup_cuda.py install

Usage

1. Reconstruction and Save Packed Model

1-1. Extract global indices for OGR

CUDA_VISIBLE_DEVICES=0 python -m qeft.extract_outidx meta-llama/Llama-2-7b-hf c4 --wbits 4 --target_rank 128 --seed 42 --no_frob_norm --output_dir global_indices/llama2-7b

1-2. Reconstruction with OGR

CUDA_VISIBLE_DEVICES=0 python -m qeft.main meta-llama/Llama-2-7b-hf c4 --wbits 4 --target_rank 128 --groupsize 128 --dtype fp16 --seed 42 --outidx_file global_indices/llama2-7b/w4_r128/outidx.pth --packing --save llama2-7b_w4_g128_r128.pth

2. Validate Packed Model Operation

2-1. Measure PPL Using Packed Model (MatMul). Result is Equal to Reconstruction

CUDA_VISIBLE_DEVICES=0 python -m qeft.main meta-llama/Llama-2-7b-hf c4 --load llama2-7b_w4_g128_r128.pth

2-2. Measure PPL Using Packed Model and Testing Acceleration(MatVec).

CUDA_VISIBLE_DEVICES=0 python -m qeft.main meta-llama/Llama-2-7b-hf c4 --benchmark 128 --load llama2-7b_w4_g128_r128.pth

# With FasterTransformer
CUDA_VISIBLE_DEVICES=0 python -m qeft.main meta-llama/Llama-2-7b-hf c4 --benchmark 128 --load llama2-7b_w4_g128_r128.pth --ft

3. Benchmark End-to-End Generation

# FP16
CUDA_VISIBLE_DEVICES=0 python -m qeft.benchmark --model_path meta-llama/Llama-2-7b-hf --method fp --ft

# QEFT 4bit
CUDA_VISIBLE_DEVICES=0 python -m qeft.benchmark --model_path meta-llama/Llama-2-7b-hf --method qeft --load ckpt/llama2-7b_w4_g128_r128.pth --ft

Reference

This code is based on various implementations and research papers related to weight quantization and model optimization.

OWQ: Outlier-Aware Weight Quantization for Efficient Fine-Tuning and Inference of Large Language Models

This code is largely based on OWQ.

AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration / Code

Platypus: Quick, Cheap, and Powerful Refinement of LLMs / Code

Cite

If you find our code useful for your research, please consider citing:

@article{lee2024qeft,
  title={QEFT: Quantization for Efficient Fine-Tuning of LLMs},
  author={Lee, Changhun and Jin, Jun-gyu and Cho, Younghyun and Park, Eunhyeok},
  journal={arXiv preprint arXiv:2410.08661},
  year={2024}
}