Not All Experts are Equal: Efficient Expert Pruning and Skipping for Mixture-of-Experts Large Language Models
Official Pytorch implementation of the expert pruning and dynamic skipping methods as presented in:
Not All Experts are Equal: Efficient Expert Pruning and Skipping for Mixture-of-Experts Large Language Models (ACL 2024, main)
Xudong Lu*,Qi Liu*, Yuhui Xu, Aojun Zhou, Siyuan Huang, Bo Zhang, Junchi Yan, Hongsheng Li (* indicates equal contribution)
CUHK MMlab, Shanghai Jiao Tong University, Shanghai Artificial Intelligence Laboratory
Paper
Step 1: Create a new conda environment:
conda create -n env_name python=3.10
conda activate env_name
Step 2: Install relevant packages
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
pip install transformers==4.36.2 accelerate datasets fire tqdm
- C4: Please download first part of the C4 training data
c4-train.00000-of-01024.json
from allenai/c4. - MATH: You can use our pre-built calibration set in
./data/math_pretrain_style.json
. To reproduce our construction, please download the training set of MATH and use our script. - Alpaca: In our project, generation speed is benchmarked using Alpaca dataset. Please download
alpaca_data_cleaned.json
from yahma/alpaca-cleaned. - Finally, please organize the calibration datasets as follows.
./data
|-- __init__.py
|-- alpaca_data_cleaned.json
|-- build.py
|-- c4-train.00000-of-01024.json
|-- dataset.py
|-- math_calib_construction.py
`-- math_pretrain_style.json
Usage:
python main.py [-h] --method {layerwise_pruning,progressive_pruning,dynamic_skipping} [--r R] --calib_set {c4,math} --model_path MODEL_PATH [--output_path OUTPUT_PATH] [--max_block_size MAX_BLOCK_SIZE] [--n_blocks_for_stat N_BLOCKS_FOR_STAT] [--batch_size BATCH_SIZE] [--num_workers NUM_WORKERS] [--seed SEED] [--use_flash_attention_2]
Options:
-h, --help
: show this help message and exit--method {layerwise_pruning,progressive_pruning,dynamic_skipping}
: Supported pruning methods: Layerwise Pruning, Progressive Pruning, Dynamic Skipping--r R
: Number of experts to preserve--calib_set {c4,math}
: Supported calibration datasets: C4, MATH--model_path MODEL_PATH
: Path to model to prune--output_path OUTPUT_PATH
: Output path (pruned model, pruning results, etc.)--max_block_size MAX_BLOCK_SIZE
: Maximal sequence length of each sample in calibration set--n_blocks_for_stat N_BLOCKS_FOR_STAT
: Number of sequences in calibration set. If set to 0 or negative, the whole dataset will be used--batch_size BATCH_SIZE
: Batch size for model inference--num_workers NUM_WORKERS
: Number of workers in dataloader--seed SEED
: Random seed for reproduction--use_flash_attention_2
: If set, Flash Attention 2 will be used
You can can perform expert pruning by running:
python main.py --method layerwise_pruning --r 6 --calib_set c4 --model_path Mixtral-8x7B-v0.1 --output_path ./output/
Note: To integrate dynamic skipping with other methods, please modify your installed transformers/models/mixtral/modeling_mixtral.py
as model/modeling_mixtral.py
We use the EleutherAI LM Harness (commit 2a47159) framework to evaluate the performance of pruned LLMs. The command we use for LM Harness evaluation is as follows:
accelerate launch \
--num_processes=1 \
--mixed_precision=bf16 \
--ipex \
-m lm_eval --model hf \
--model_args pretrained=model_path,dtype=bfloat16,parallelize=True \
--tasks boolq,rte,arc_challenge,arc_easy,hellaswag,winogrande,openbookqa,gsm8k,mmlu\
--batch_size 16
We use eval/benchmark_speed.py to evaluate the speedup of pruned LLMs. The command we use for speedup evaluation is as follows:
python ./benchmark_speed.py --num_samples 50 --model_name_or_path Mixtral-8x7B-v0.1
For finer-grained options, please refer to the script.
To benchmark generation speed for semi-structured pruned models, please add a new item wanda_sparsity_type
into model.config
(e.g., 2:4).
This project is released under the MIT license. Please see the LICENSE file for more information.
Feel free to discuss papers/code with us through issues/emails!
- Xudong Lu: luxudong@link.cuhk.edu.hk
- Qi Liu: purewhite@sjtu.edu.cn
If you find our paper and code useful in your research, please cite
@misc{lu2024experts,
title={Not All Experts are Equal: Efficient Expert Pruning and Skipping for Mixture-of-Experts Large Language Models},
author={Xudong Lu and Qi Liu and Yuhui Xu and Aojun Zhou and Siyuan Huang and Bo Zhang and Junchi Yan and Hongsheng Li},
year={2024},
eprint={2402.14800},
archivePrefix={arXiv},
primaryClass={cs.CL}
}