Skip to content

Latest commit

 

History

History
123 lines (89 loc) · 4.46 KB

README.md

File metadata and controls

123 lines (89 loc) · 4.46 KB
Raw

ComPEFT: Compression for Communicating Parameter Efficient Updates via Sparsification and Quantization

teaser image

Setup

  1. Create a virtual environment and activate it.
conda create -n compeft python==3.9
conda activate compeft
pip install -r requirements.txt
cd src
  1. Set the path to where finetuned models are stored in src/utils.py

Train

Finetune BERT like Models.

To train checkpoint for BERT like models you can use the file src/bert_finetune.py. A sample command for training MRPC from GLUE is given below.

Full Model finetuning

python bert_finetune.py --learning_rate=5e-05 --train --test --model_name_or_path=bert-large-uncased --num_epochs=10 --project_name=ft_bertlarge --task=mrpc

(IA)^3 Finetuning

python bert_finetune.py --learning_rate=0.005 --peft=ia3 --train --test --model_name_or_path=bert-large-uncased --num_epochs=10 --project_name=ft_bertlarge --task=mrpc

LoRA Finetuning

python bert_finetune.py --learning_rate=0.0005 --peft=lora --train --test --model_name_or_path=bert-large-uncased --num_epochs=10 --project_name=ft_bertlarge --task=mrpc

Finetune T5 like models.

To train checkpoint for T5 like models (T5, T5v1.1, T0) you can use the file src/t5_finetune.py. A sample command for training RTE from GLUE is given below.

Full Model finetuning

python t5_finetune.py --project_name=ft_t5base --model_name_or_path=t5-base --train=True --task=rte --num_epochs=2 --learning_rate=3e-4

(IA)^3 Finetuning

python t5_finetune.py --project_name=ft_t5base --model_name_or_path=t5-base --train=True --task=rte --num_epochs=2 --learning_rate=3e-4 --peft=ia3

LoRA Finetuning

python t5_finetune.py --project_name=ft_t5base --model_name_or_path=t5-base --train=True --task=rte --num_epochs=2 --learning_rate=3e-4 --peft=lora

Compress using Compeft and evaluate.

$K = # set density between 0-100

$replace_factor = # set the values of alpha.

BERT-like models

Compress Fully Finetuned BERT Model.

python bert_eval.py --model_name_or_path t5-base --task rte  --k ${K} --replace_factor ${replace_factor} --project_name eval_t5base

Compress (IA)^3 on BERT Model.

python bert_eval.py --model_name_or_path t5-base --task rte --peft ia3 --k ${K} --replace_factor ${replace_factor} --project_name eval_t5base

Compress LoRA on BERT Model.

python bert_eval.py --model_name_or_path t5-base --task rte --peft lora  --k ${K} --replace_factor ${replace_factor} --project_name eval_t5base

T5-like models

Compress Fully Finetuned T5 Model.

python t5_eval.py --model_name_or_path t5-base --task rte --k ${K} --replace_factor ${replace_factor} --project_name eval_t5base

Compress (IA)^3 on T5 Model.

python t5_eval.py --model_name_or_path t5-base --task rte --peft ia3 --k ${K} --replace_factor ${replace_factor} --project_name eval_t5base

Compress LoRA on T5 Model.

python t5_eval.py --model_name_or_path t5-base --task rte --peft lora  --k ${K} --replace_factor ${replace_factor} --project_name eval_t5base

Compress QLoRA on LLaMA

$model_name_or_path = # LLaMA checkpoint dir.

$checkpoint_dir = #lora checkpoint dir

$output_dir = # output dir to store results.

python qlora_eval.py --project_name eval_llama  --model_name_or_path ${model_name_or_path} --checkpoint_dir ${checkpoint_dir} --k ${K} --replace_factor ${replace_factor} --output_dir ${output_dir} --mmlu_split eval --report_to wandb --data_seed 42 --per_device_eval_batch_size 1 --max_new_tokens 32 --dataloader_num_workers 4 --logging_strategy steps --remove_unused_columns False --group_by_length True --do_eval True --double_quant True --bf16 True --quant_type nf4 --bits 4 --dataset oasst1 --source_max_len 512 --target_max_len 512 --seed 0

Reference

Please consider citing our paper if you use our code or models in your work:

@misc{yadav2023compeft,
      title={ComPEFT: Compression for Communicating Parameter Efficient Updates via Sparsification and Quantization}, 
      author={Prateek Yadav and Leshem Choshen and Colin Raffel and Mohit Bansal},
      year={2023},
      eprint={2311.13171},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}