README.md

Prune As You Tune: LoRA-Enabled Model Compression

This repository contains the codebase for the paper "Prune As You Tune: LoRA-Enabled Model Compression," our CS2420 Final Project.

Graphical overview of our method. Prune As You Tune (PAYT) interleaves pruning of pre-trained parameters with LoRA fine-tuning updates guided by a custom Knowledge Distillation loss function. PAYT can achieve up to 50% sparsity with minimal accuracy degradation and lower perplexity on the original task compared to baselines such as full fine-tuning and prune-then-fine-tune methods. Compared to state-of-the-art approaches, PAYT offers superior accuracy, model compression, and retention of pre-training knowledge.

src contains the Prune As You Tune implementation.
dev contains initial experimentation and tests.
plot contains plotting code samples.

src/main.py contains our typer app for running experiments.

Setup

Create an environment and install dependencies with

pip install -r requirements.txt

Usage

Easily evaluate various Prune As You Tune configurations and compare against baseline methods with our typer app.

First, execute

export PYTHONPATH=`pwd`

Run the typer app with

typer main.py run [OPTIONS]

The table below gives the options supported by our app. Run typer main.py run --help to reproduce it.

Option	Type	Description	Default Value
--model-name	TEXT	Model name to use for fine-tuning	bert-base-uncased
--n-samples	INTEGER	Number of samples, use 10 or less for rapid testing	1000
--sparsity-target	FLOAT	Target percentage of parameters to prune	0.5
--num-epochs	INTEGER	Number of training epochs	5
--output-dir	TEXT	Output directory for logs and model checkpoints	logs
--dataset	TEXT	Dataset to use for fine-tuning	imdb
--dataset-mix-ptg	FLOAT	Percentage of original dataset samples to include - in training only	0.05
--dataset-mix-strategy	TEXT	Mixing strategy for dataset - old_first or random	random
--dataset-sampling-strategy	TEXT	Sampling strategy for old samples dataset - first or random	random
--full-evaluate, --no-full-evaluate		Evaluate using all variations of pruning methods, as well as full fine-tuning	no-full-evaluate
--use-lora, --no-use-lora		Use LoRA adapters for fine-tuning	use-lora
--use-kd, --no-use-kd		Use knowledge distillation for fine-tuning	use-kd
--kd-alpha	FLOAT	Alpha parameter for knowledge distillation	0.5
--kd-temp	FLOAT	Temperature parameter for knowledge distillation	4
--use-rs-lora, --no-use-rs-lora		Use rsLoRA adapters for fine-tuning	use-rs-lora
--lora-dropout	FLOAT	Dropout rate for LoRA adapters	0.1
--lora-alpha	FLOAT	Scaling factor for LoRA adapters	32
--lora-rank	INTEGER	Rank of LoRA adapters	32
--max-length	INTEGER	Maximum length of input sequences	512
--pruning-schedule	TEXT	Pruning schedule - can be agp or linear	linear
--prune-every-epoch	INTEGER	Prune at every n epoch	1
--start-pruning-epoch-ptg	FLOAT	Start pruning at n percent epoch of training - will be rounded down	0
--pruning-method	TEXT	Pruning method - can be L1Unstructured or L2Structured	L1Unstructured
--help		Show this message and exit

---

Examples

# fine-tune on IMDb (default) for 10 epochs with the AGP pruning schedule and 20% delay
# use KD loss with alpha = 0.4 and T = 4, and no data mixing

typer main.py run --n-samples 25000 --pruning-schedule "agp" --num-epochs 10 --start-pruning-epoch-ptg 0.2 --kd-alpha 0.4 --kd-temp 4 --dataset-mix-ptg 0

# fine-tune on IMDb for 5 epochs with the AGP pruning schedule and no KD loss
# mix WikiText-2 into 5% of training samples

typer main.py run --n-samples 25000 --pruning-schedule "agp" --num-epochs 5 --no-use-kd --dataset-mix-ptg 0.05 

# fine-tune on SST-5 for 10 epochs with AGP, to just 30% sparsity
# use KD loss with alpha = 0.4 and T = 4, and no data mixing

typer main.py run --dataset "SetFit/sst5" --pruning-schedule "agp" --sparsity-target 0.3 --num-epochs 10 --kd-alpha 0.4 --kd-temp 4 --dataset-mix-ptg 0

# run full evaluation suite, including baseline methods

typer main.py run --full-evaluate

See src/evaluator.py for all fine-tune evaluation methods, including baselines such as full_finetune() and prune_lora_finetune().

Note that by default, perplexity on WikiText-2 is calculated and printed at the completion of each fine-tuning run. (See eval_ppl=True in main.py.)

Checkpoints and logs are saved at the end of epoch.
Sparsity and perplexity data are printed to standard output, in addition to other training information.

For reference: Training on IMDb for 5 epochs with KD (followed by perplexity evaluation) took us about 60 minutes on one A100 GPU.

Authors

• Joey Bejjani @jbejjani2022
• Amulya Garimella @amulyagarimella
• Swati Goel @sgGoel
• Mohammed Sakib @MohammedSakib7