LLM-Pruner

On the Structural Pruning of Large Language Models

🦙 🦙 🦙 🦙 🦙 Compress your LLMs to any size! 🦙 🦙 🦙 🦙 🦙

LLM-Pruner: On the Structural Pruning of Large Language Models [arXiv]
Xinyin Ma, Gongfan Fang, Xinchao Wang
National University of Singapore

• Task-agnostic compression: The compressed LLM should retain its original ability as a multi-task solver.
• Less training corpus: In this work, we use only 50k publicly available samples (alpaca) to post-train the LLM.
• Efficient compression: 3 minutes for pruning and 3 hours for post-training. (You can make it longer)
• Automatic structural pruning: Pruning new LLMs with minimal human effort (In progress).

• Llama-3.1
• Llama-3
• Llama-2
• LLaMA
• BLOOM
• Vicuna
• Baichuan
• TinyLlama

• July 27, 2024: 🚀 Support GQA! Now LLM-Pruner can work on Llama3 and Llama 3.1. We are still testing the pruning results of new LLMs (Llama3, Llama3.1, Gemma) and you can find the pruning results here.
• August 30, 2023: LLM-Pruner now supports BLOOM 🌸
• August 14, 2023: Code and results for finetuning with a large-scale corpus are now available. The fine-tuned LLaMA-5.4B model achieves an average accuracy of 62.36%, closely approaching the original LLaMA-7B (63.25%).
• July 19, 2023: 🔥 LLM-Pruner now supports Llama-2-7b and Llama-2-13b (the huggingface version)
• July 18, 2023: 🚀 Support Baichuan, a bilingual LLM.
• May 20, 2023: 🎉 Code and Preprint Paper released!

• A tutorial for pruning new LLMs.
• Support .from_pretrained() for loading the model.

Join our WeChat group for a chat:

• WeChat Group Group-2 (>200/500), Group-1 (500/500, FULL).

• Quick Start
• Step-by-step Instructions
• Zero-shot Evaluation
• More-Examples
• Version Information
• Limitations
• Acknowledgement
• Citation

pip install -r requirement.txt

bash script/llama_prune.sh

This script would compress the LLaMA-7B model with ～20% parameters pruned. All the pre-trained models and the dataset would be automatically downloaded, so you do not need to manually download the resource. When running this script for the first time, it will require some time to download the model and the dataset.

It takes three steps to prune an LLM:

• Discovery Stage: Discover the complicated inter-dependency in LLMs and find the minimally-removable unit, group.
• Estimation Stage: Estimate the contribution of each group to the overall performance of the model and decide which group to prune.
• Recover Stage: Fast post-training to recover model performance.

After pruning and post-training, we follow lm-evaluation-harness for evaluation.

🦙 LLaMA/Llama-2 pruning with ~20% parameters pruned:

python hf_prune.py --pruning_ratio 0.25 \
      --block_wise \
      --block_mlp_layer_start 4 --block_mlp_layer_end 30 \
      --block_attention_layer_start 4 --block_attention_layer_end 30 \
      --pruner_type taylor \
      --test_after_train \
      --device cpu  --eval_device cuda \
      --save_ckpt_log_name llama_prune 

Arguments:

• Base model: Choose the base model from LLaMA or Llama-2 and pass the pretrained_model_name_or_path to --base_model. The model name is used for AutoModel.from_pretrained to load the pre-trained LLM. For example, if you want to use the llama-2 with 13 billion parameters, then pass meta-llama/Llama-2-13b-hf to --base_model.
• Pruning Strategy: Choose between block-wise, channel-wise, or layer-wise pruning using the respective command options: {--block_wise}, {--channel_wise}, {--layer_wise --layer NUMBER_OF_LAYERS}. For block-wise pruning, specify the start and end layers to be pruned. Channel-wise pruning does not require extra arguments. For layer pruning, use --layer NUMBER_OF_LAYERS to specify the desired number of layers to be kept after pruning.
• Importance Criterion: Select from l1, l2, random, or taylor using the --pruner_type argument. For the taylor pruner, choose one of the following options: vectorize, param_second, param_first, param_mix. By default, param_mix is used, which combines approximated second-order hessian and first-order gradient. If using l1, l2, or random, no extra arguments are required.
• Pruning Ratio: Specifies the pruning ratio of groups. It differs from the pruning rate of parameters, as groups are removed as the minimal units.
• Device and Eval_device: Pruning and evaluation can be performed on different devices. Taylor-based methods require backward computation during pruning, which may require significant GPU RAM. Our implementation uses the CPU for importance estimation (also supports GPU, simply use --device cuda). eval_device is used to test the pruned model.

python hf_prune.py --pruning_ratio 0.25 \
      --block_wise \
      --block_mlp_layer_start 4 --block_mlp_layer_end 30 \
      --block_attention_layer_start 4 --block_attention_layer_end 30 \
      --pruner_type taylor \
      --test_after_train \
      --device cpu  --eval_device cuda \
      --save_ckpt_log_name llama_prune \
      --base_model PATH_TO_VICUNA_WEIGHTS

python llama3.py --pruning_ratio 0.25 \
                 --device cuda --eval_device cuda \
                 --base_model meta-llama/Meta-Llama-3-8B-Instruct \
                 --block_wise --block_mlp_layer_start 4 --block_mlp_layer_end 30 \
                 --block_attention_layer_start 4 --block_attention_layer_end 30 \
                 --save_ckpt_log_name llama3_prune \
                 --pruner_type taylor --taylor param_first \
                 --max_seq_len 2048 \
                 --test_after_train --test_before_train --save_model 

• Train using Alpaca with 50,000 samples. Here's an example of training on a single GPU:

CUDA_VISIBLE_DEVICES=X python post_training.py --prune_model prune_log/PATH_TO_PRUNE_MODEL/pytorch_model.bin \
      --data_path yahma/alpaca-cleaned \
      --lora_r 8 \
      --num_epochs 2 \ 
      --learning_rate 1e-4 \ 
      --batch_size 64 \
      --output_dir tune_log/PATH_TO_SAVE_TUNE_MODEL \ 
      --wandb_project llama_tune

Make sure to replace PATH_TO_PRUNE_MODEL with the path to the pruned model in step 1, and replace PATH_TO_SAVE_TUNE_MODEL with the desired location where you want to save the tuned model.

Tip: Training LLaMA-2 in float16 is not recommended and is known to produce nan; as such, the model should be trained in bfloat16.

• Train using MBZUAI/LaMini-instruction with 2.59M samples. Here is an example using multiple gpus for training:

deepspeed --include=localhost:1,2,3,4 post_training.py \
      --prune_model prune_log/PATH_TO_PRUNE_MODEL/pytorch_model.bin \
      --data_path MBZUAI/LaMini-instruction  \
      --lora_r 8 \
      --num_epochs 3  \
      --output_dir tune_log/PATH_TO_SAVE_TUNE_MODEL \
      --extra_val_dataset wikitext2,ptb \
      --wandb_project llmpruner_lamini_tune \
      --learning_rate 5e-5 \
      --cache_dataset

For the pruned model, simply use the following command to load your model.

  pruned_dict = torch.load(YOUR_CHECKPOINT_PATH, map_location='cpu')
  tokenizer, model = pruned_dict['tokenizer'], pruned_dict['model']

Due to the different configurations between modules in the pruned model, where certain layers may have larger width while others have undergone more pruning, it becomes impractical to load the model using the .from_pretrained() as provided by Hugging Face. Currently, we employ the torch.save to store the pruned model.

Since the pruned model has different configuration in each layer, like some layers might be wider but some layers have been pruned more, the model cannot be loaded with the .from_pretrained() in Hugging Face. Currently, we simply use the torch.save to save the pruned model and torch.load to load the pruned model.

We provide a simple script to geneate texts using pre-trained / pruned models / pruned models with post-training.

• LLaMA-7B Pre-trained

python generate.py --model_type pretrain

• Pruned Model without Post-Training

python generate.py --model_type pruneLLM --ckpt <YOUR_MODEL_PATH_FOR_PRUNE_MODEL>

• Pruned Model with Post-Training

python generate.py --model_type tune_prune_LLM --ckpt <YOUR_CKPT_PATH_FOR_PRUNE_MODEL> --lora_ckpt <YOUR_CKPT_PATH_FOR_LORA_WEIGHT>

The above instructions will deploy your LLMs locally.

For evaluating the performance of the pruned model, we follow lm-evaluation-harness to evaluate the model:

• Step 1: If you only need to evaluate the pruned model, then skip this step and jump to Step 2. This step is to arrange the files to satisfy the input requirement for lm-evaluation-harness. The tuned checkpoint from the post-training step would be save in the following format:

- PATH_TO_SAVE_TUNE_MODEL
  | - checkpoint-200
      | - pytorch_model.bin
      | - optimizer.pt
      ...
  | - checkpoint-400
  | - checkpoint-600
  ...
  | - adapter_config.bin
  | - adapter-config.json

Arrange the files by the following commands:

cd PATH_TO_SAVE_TUNE_MODEL
export epoch=YOUR_EVALUATE_EPOCH
cp adapter_config.json checkpoint-$epoch/
mv checkpoint-$epoch/pytorch_model.bin checkpoint-$epoch/adapter_model.bin

If you want to evaluate the checkpoint-200, then set the epoch equalts to 200 by export epoch=200.

• Step 2:

export PYTHONPATH='.'
python lm-evaluation-harness/main.py --model hf-causal-experimental \
       --model_args checkpoint=PATH_TO_PRUNE_MODEL,peft=PATH_TO_SAVE_TUNE_MODEL,config_pretrained=PATH_OR_NAME_TO_BASE_MODEL \
       --tasks openbookqa,arc_easy,winogrande,hellaswag,arc_challenge,piqa,boolq \
       --device cuda:0 --no_cache \
       --output_path PATH_TO_SAVE_EVALUATION_LOG 

Here, replace PATH_TO_PRUNE_MODEL and PATH_TO_SAVE_TUNE_MODEL with the path you save the pruned model and the tuned model, and PATH_OR_NAME_TO_BASE_MODEL is for loading the configuration file of the base model.

[Update]: We upload a script to simply the evaluation process if you want to evaluate the pruned model with the tuned checkpoint. Simply use the following command:

CUDA_VISIBLE_DEVICES=X bash scripts/evaluate.sh PATH_OR_NAME_TO_BASE_MODEL PATH_TO_SAVE_TUNE_MODEL  PATH_TO_PRUNE_MODEL EPOCHS_YOU_WANT_TO_EVALUATE

Replace the necessary information of your model in the command. The final one is used to iterate over different epochs if you want to evaluate several checkpoints in one command. For example:

CUDA_VISIBLE_DEVICES=1 bash scripts/evaluate.sh decapoda-research/llama-7b-hf tune_log/llama_7B_hessian prune_log/llama_prune_7B 200 1000 2000

• Pre-trained

python test_speedup.py --model_type pretrain

• Pruned Model

python test_speedup.py --model_type pruneLLM --ckpt <YOUR_MODEL_PATH_FOR_PRUNE_MODEL>

A brief quantitative results for LLaMA-7B:

The results for Vicuna-7B:

The results for ChatGLM-6B:

Statistics for pruned models:

Results of LLM-Pruner with 2.59M samples:

More results can be found in the paper.

Due to changes in the versions of models and repos used in this project, we listed some known version issues and the specific versions needed to reproduce our method:

1. lm-eval-harness: We use this commit of lm-evaluation-harness, and the code is also included in this repo. Please check Issue #25 for details.
2. LLaMA1-7B: We use the checkpoint of decapoda-research/llama-7b-hf in our experiments, which is not available now. Please consider using the copied version, e.g.,baffo32/decapoda-research-llama-7B-hf.

• Although we only used 50K data and trained for three hours, more data would definitely be better. We are testing on this.
• The current compressed model still has several issues, such as generating repetitive tokens or producing nonsensical sentences. We believe there is significant room for improvement in the quality of the compressed model.
• There are still some models for which we cannot automatically identify the mapping of indexes after concatenation and view operations. Therefore, we need to perform additional manual operations.

• Logo is generated by Stable Diffusion
• The evaluation of the LLM: lm-evaluation-harness
• LLaMA: https://github.com/facebookresearch/llama
• Vicuna: https://github.com/lm-sys/FastChat
• Peft: https://github.com/huggingface/peft
• Alpaca-lora: https://github.com/tloen/alpaca-lora

If you find this project useful, please cite

@inproceedings{ma2023llmpruner,
  title={LLM-Pruner: On the Structural Pruning of Large Language Models},
  author={Xinyin Ma and Gongfan Fang and Xinchao Wang},
  booktitle={Advances in Neural Information Processing Systems},
  year={2023},
}

@article{fang2023depgraph,
  title={DepGraph: Towards Any Structural Pruning},
  author={Fang, Gongfan and Ma, Xinyin and Song, Mingli and Mi, Michael Bi and Wang, Xinchao},
  journal={The IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2023}
}