LoftQ: LoRA-Fine-Tuning-Aware Quantization

LoftQ helps you fine-tune LLMs with limited GPUs. 🚀 LoftQ finds good enough quantized LoRA initialization: quantized backbone Q and LoRA adapters A and B, given a pre-trained weight W.

This repo implements the paper 🔗: LoftQ: LoRA-Fine-Tuning-Aware Quantization for Large Language Models.

Our models are available on 🤗 LoftQ Huggingface Hub

News

• [04/20/2024] New LLAMA-3-8Bresults on GSM8K. See results here. Check out LoftQ of 🦙 LLAMA-3, CodeLLAMA-7b, CodeLLAMA=13b on Huggingface Hub.

• [04/13/2024] New phi-2 results on GSM8K. See results here. Check out LoftQ of Phi-2 on Huggingface Hub.

• [04/13/2024] Update script/train_gsm8k.sh to support data parallel of quantized models.

Quick Start

Requirements

We use bitsandbytes to implement the quantization. This package only support CUDA >= 11.0 and does not support CPU. However, we also provide fake quantization for fast and parallel training if GPUs are adequate.

pip install -r requirements.txt

Steps

1. Apply LoftQ to a full-precision pre-trained weight and save.
2. Load LoftQ initialization and train.

For step 1, we have provided off-the-shelf LoftQ initializations (see supported model list) in Huggingface Hub LoftQ. If you want to do it yourself, jump to LoftQ DIY.

For step 2, below is an example of loading 4bit Mistral-7B with 64rank LoRA adapters from Huggingface Hub.

import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

# fetch the MODEL_ID at https://huggingface.co/LoftQ
MODEL_ID = "LoftQ/Mistral-7B-v0.1-4bit-64rank"

base_model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, 
    torch_dtype=torch.bfloat16,  # you may change it with different models
    quantization_config=BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.bfloat16,  # bfloat16 is recommended
        bnb_4bit_use_double_quant=False,
        bnb_4bit_quant_type='nf4',
    ),
)
peft_model = PeftModel.from_pretrained(
    base_model,
    MODEL_ID,
    subfolder="loftq_init",
    is_trainable=True,
)

# Do training with peft_model ...

LoftQ DIY

Apply LoftQ and save

We provide quantize_save.py as an example to apply LoftQ with different bits(--bits), ranks(--rank), and alternating steps (--iter, a hyper-parameter in LoftQ, see Algorithm 1 in LoftQ paper). Currently, this example supports llama-2, falcon, mistral, bart, t5, deberta, bert, roberta.

Below is an example of obtaining 4bit LLAMA-2-7b with 16-rank LoRA adapters by 5 alternating steps.

SAVE_DIR="model_zoo/loftq/"
python quantize_save_load.py \
    --model_name_or_path meta-llama/Llama-2-7b-hf \  # high-precision model id in HF
    --token HF_TOKEN \  # your HF token if the model is private, e.g., llama-2
    --bits 4 \
    --iter 5 \
    --rank 16 \
    --save_dir $SAVE_DIR

The above commands end up with creating the model directory under $SAVE_DIR. Specifically, the model directory is named as

MODEL_DIR = SAVE_DIR + f"{args.model_name_or_path.split('/')[-1]}-{args.bits}bits-{args.rank}rank"

In this example, MODEL_DIR="model_zoo/loftq/Llama-2-7b-hf-4bit-16rank", where the backbone is stored in $MODEL_DIR and the LoRA adapters are at the sub-folder $MODEL_DIR/loftq_init.

Load and train

Similar to loading from Huggingface Hub, we only need to change the MODEL_ID to the MODEL_DIR.

import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

MODEL_DIR = "model_zoo/loftq/Llama-2-7b-hf-4bit-16rank"

base_model = AutoModelForCausalLM.from_pretrained(
    MODEL_DIR, 
    torch_dtype=torch.bfloat16,
    quantization_config=BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.bfloat16,
        bnb_4bit_use_double_quant=False,
        bnb_4bit_quant_type='nf4',
    ),
)
peft_model = PeftModel.from_pretrained(
    base_model,
    MODEL_DIR,
    subfolder="loftq_init",
    is_trainable=True,
)
# Do training with peft_model ...

LoftQ Fine-tuning

We also provide an example to fine-tune LLAMA-7b with LoftQ on GSM8K.

python train_gsm8k.py \
    --model_name_or_path LoftQ/Llama-2-7b-hf-4bit-64rank \
    --learning_rate 3e-4 \
    --seed 11 \
    --expt_name gsm8k_llama2_7b_4bit_64rank_loftq \
    --output_dir exp_results/ \
    --num_train_epochs 6 \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "no" \
    --save_strategy "epoch" \
    --weight_decay 0.1 \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 10 \
    --do_train \
    --report_to tensorboard

Other training Files

• GLUE: glue/run_glue.py
• Question Answering: glue/run_qa.py
• Summarization: train_summarization.py
• WikiText-2: train_clm.py
• GSM8K: train_gsm8k.py

More example scripts are in scripts.

Quick Evaluation

Here is the command to test GSM8K with adapters we have fine-tuned. It is stored in the subfolder='gsm8k' of the target model in LoftQ Huggingface hub.

python test_gsm8k.py \
    --model_name_or_path LoftQ/Llama-2-7b-hf-4bit-64rank \
    --batch_size 16

python test_gsm8k.py \
    --model_name_or_path LoftQ/phi-2-4bit-64rank \
    --batch_size 16

Feel free to change batch_size to accommodate to your machine.

Main Results

LLAMA-2 on WikiText-2 and GSM8K

Models are fine-tuned through causal language modeling on training sets and are tested on validation/test sets.

Phi-2 on GSM8K

LLAMA-3 on GSM8K

Models are fine-tuned through causal language modeling on (reformatted) training sets and are tested on validation/test sets.

BART-large on CNN/DailyMail and XSum

*: Using the first 3 sentences in the document as the summary

DeBERTa-V3-base on GLUE using Normal Float Datatype

DeBERTa-V3-base on GLUE using Uniform Quantization Datatype

Citation

@article{li2023loftq,
  title={Loftq: Lora-fine-tuning-aware quantization for large language models},
  author={Li, Yixiao and Yu, Yifan and Liang, Chen and He, Pengcheng and Karampatziakis, Nikos and Chen, Weizhu and Zhao, Tuo},
  journal={arXiv preprint arXiv:2310.08659},
  year={2023}
}

Appendix: Off-the-shelf Model List