[ACL 2024] AFLoRA: Adaptive Freezing of Low Rank Adaptation in Parameter Efficient Fine-Tuning of Large Models

This is the official repository of AFLoRA, accepted in ACL 2024.

Schematic comparison of LoRA, ELoRA (VeRA), and AFLoRA

Contributors

1. Zeyu Liu* (USC)
2. Souvik Kundu* (Intel Labs)

Directory Structure

glue

• Modified code for running experiments on the GLUE benchmark.
• Adapted from LoRA source code and VeRA.
• We define the CustomTrainerWFreeze in glue/trainer.py to adptively freeze the projection matrices in the LoRA path.
• The details of how we choose a threshold to freeze the projection matrices are shown in glue/utils.py.

peft

• Modified Hugging Face PEFT library.
• The only two relevant files for our work is peft/src/peft/tuners/_lora.py and peft/src/peft/tuners/lora.py. We reproduce VeRA based on their code and add our method upon that.
• The rest of the library is unchanged.

gsm8k

• Modified code for running experiments on the GSM8K dataset.
• Adapted from LLM-Adapters.
• We define the CustomTrainerWFreeze in gsm8k/trainer.py to adptively freeze the projection matrices in the LoRA path.
• The details of how we choose a threshold to freeze the projection matrices are shown in gsm8k/utils.py.

peft

• Modified Hugging Face PEFT library.
• The only two relevant files for our work is peft/src/peft/tuners/lora.py. We reproduce VeRA based on their code and add our method upon that.
• The rest of the library is unchanged.

Requirements (The requirements are different for the task GLUE and GSM8K)

• Python 3.10
• Run pip install -r requirements.txt to install necessary packages.
• Run pip install -e ./peft to install modified peft library.

Main arguments

# For GLUE
export CUDA_VISIBLE_DEVICES='0' 
CUDA_DEVICE=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f1)
BASE_PORT=29510
MASTER_PORT=$(($BASE_PORT + $CUDA_DEVICE)) 
torchrun --nproc_per_node=1 --master_port $MASTER_PORT run_glue.py \
         --wandb_offline 0 \ decide if use wandb in offline mode
         --do_train \
         --do_eval \
         --gradient_accumulation_steps 1 \
         --output_dir ./output \
         --overwrite_output_dir \
         --logging_steps 10 \
         --logging_dir ./output/log \
         --evaluation_strategy epoch \
         --save_strategy epoch \
         --warmup_ratio 0.06 \
         --max_grad_norm 0.1 \ gradient clip
         --weight_decay 0.1 \
         --shared_uv 0 \
         --model_name_or_path microsoft/deberta-v3-base \
         --tokenizer_name microsoft/deberta-v3-base \
         --per_device_train_batch_size 64 \
         --max_seq_length 256 \
         --mode elora \
         --lora_r 4 \ rank of LoRA
         --init_type 1 \
         --d_init_type 94 \
         --seed 42 \
         --task_name mrpc \ {mrpc, stsb, cola, ...}
         --num_train_epochs 30 \
         --classifier_lr 8e-2 \
         --learning_rate 1e-2 \
         --trainable_uv 1 \ if train the lora_A and lora_B matrices
         --disable_tqdm true \
         --freeze_by_epoch 0 \ after xxx epochs, freeze the lora_A and lora_B
         --freeze_by_ipt false \ if freeze the lora_A and lora_B based on their important scores
         --lora_dropout 0.0 \
         --load_best_model_at_end true \

# For GSM8K
export CUDA_VISIBLE_DEVICES='0' 
CUDA_DEVICE=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f1)
BASE_PORT=29520
MASTER_PORT=$(($BASE_PORT + $CUDA_DEVICE))  
WORLD_SIZE=1
python finetune.py \
  --base_model 'yahma/llama-7b-hf' \
  --data_path 'ft-training_set/math_10k.json' \
  --output_dir './trained_models/llama-7b-math10k/' \
  --batch_size 1  --micro_batch_size 1 \
  --num_epochs 3   --learning_rate 2e-3 \
  --cutoff_len 256   --val_set_size 120 \
  --eval_step 80 --save_step 80  \
  --adapter_name aflora \
  --target_modules '["q_proj", "k_proj", "v_proj", "up_proj", "down_proj"]' \
  --lora_r 32 --lora_alpha 64 \
  --wandb_project 'llama-7b-math10k'

Main Results

Comparison of different LoRA variants with DeBERTaV3 on the GLUE benchmark.

Method	#Params. ↓	CoLA ↑	SST-2 ↑	MRPC ↑	QNLI ↑	STS-B ↑	RTE ↑	MNLI ↑	QQP ↑	Avg. ↑
FFT	184M	69.21	95.64	89.22	93.78	91.59	82.49	89.98/89.95	92.05/89.31	87.82
LoRA (r = 8)	1.33M	69.73	95.57	89.71	93.76	91.86	85.32	90.47/90.46	91.95/89.26	88.38
AdaLoRA	1.27M	70.86	95.95	90.22	94.28	91.39	87.36	90.27/90.30	92.13/88.41	88.83
SoRA (r = 4)	0.47M	71.05	95.57	90.20	93.92	91.76	86.04	90.38/90.43	92.06/89.44	88.71
ELoRA*	0.16M	70.74	95.18	90.93	93.58	91.08	87.36	90.11/90.22	90.69/87.63	88.53
AFLoRA (r = 4)	0.14M	72.01	96.22	91.91	94.42	91.84	88.09	89.88/90.17	90.81/87.77	89.23

*The original paper has results with the RoBERTa, we generated the results with our implementation on DeBERTaV3 with rank of 1024.
**As the number of trainable parameters is changed during training, we computed this by averaging over the whole training epochs over all datasets.

Visualization of freezing iterations for each layer.

‘out’ and ‘inter’ refer to the second and the first MLP layer of the FFN, respectively. ‘A’ and ‘B’ represent the down-projection and up-projection matrix, respectively. The darker the color, the more iterations the matrix has to go through before freezing.

Citing

Please consider citing our paper if you find this repository useful.

@misc{liu2024aflora,
      title={AFLoRA: Adaptive Freezing of Low Rank Adaptation in Parameter Efficient Fine-Tuning of Large Models}, 
      author={Zeyu Liu and Souvik Kundu and Anni Li and Junrui Wan and Lianghao Jiang and Peter Anthony Beerel},
      year={2024},
      journal={ACL 2024},
}

Acknowledgments

The code for this project references the following previous work:

PEFT

VeRA (ELoRA)

LoSparse

LLM-Adapters.