[ACL 2024] BitDistiller: Unleashing the Potential of Sub-4-Bit LLMs via Self-Distillation [paper]

Implementing efficient sub-4-bit weight quantization (3 / 2 bits) in LLMs through advanced QAT-based Self-Distillation techniques.

Comparing general language tasks with other methods

Comparing reasoning benchmarks with other methods

Example on 2-bit inference of a Domain-specific LLM (MetaMath)

News

• [2024/05] 🔥 BitDistiller has been accepted to ACL main 2024!

Contents

1. Setup
2. Running
3. Evaluation
4. Inferencce

1. Setup

• python 3.9, pytorch >= 1.13
• pip install -r requirement.txt

(You may need to change the version of transformers according to the model config)

2. Running

Our results is running by following 3 steps:

2.1. Asymmetric Quantization

• Determine the type of quantization: use nf3 for 3 bits and int for 2 bits. Set w_bit and quant_type accordingly.

• Perform clipping before training and save the clipping values using dump_clip (see quantization/autoclip.py).

This step can match or surpass the low-bit PTQ quantization results of GPTQ and AWQ.

2.2. Generating Teacher Data

• For QAT, create data using the Teacher Model (BF16). The data varies depending on the model (see data/generation).

2.3. KD-base QAT

• Detailed procedure available in train/

Example Srcipts

LLaMA-2

1. Get the Clipping result ```bash cd BitDistiller/quantization

   CUDA_VISIBLE_DEVICES=0 python autoclip.py --model_path --calib_dataset pile --quant_type int --w_bit 2 --q_group_size 128 --run_clip --dump_clip ./clip_cache/hf-llama2-7b/int2-g128.pt 2. Get the Teacher Generation Data (Using vllm would be much faster)bash

   vllm

   python generate_vllm.py --base_model --dataset_name wikitext --out_path ./datasets/hf-llama-2-7b/ --max_sample 3000

   python generate_vllm.py --base_model --dataset_name alpaca --out_path ./datasets/hf-llama-2-7b/ --max_sample 5000

   change to path in .py

   python mix_data.py ```

   ```bash

   torchrun

   cd BitDistiller/data/generation

   bash generate.sh wikitext ../datasets/hf-llama-2-7b/ 16 3000

   bash generate.sh alpaca ../datasets/hf-llama-2-7b/ 16 5000

   change to path in .py

   python mix_data.py 3. Run KD-base QATbash

   Specify the pre-trained model path

   Specify the num_gpus and batch_size according to your GPU devices

   Specify the clipping cache path to the --clip

   cd train

   bash train.sh ../data/datasets/hf-llama-2-7b/mix_wiki_alpaca_8000.json ./ckpts/hf-llama-2-7b/int2-g128/ ./logs/hf-llama-2-7b/int2-g128/ 4 ```

WizardCoder

1. Get the Clipping result ```bash cd BitDistiller/quantization

   CUDA_VISIBLE_DEVICES=0 python autoclip.py --model_path --calib_dataset code --quant_type int --w_bit 2 --q_group_size 128 --run_clip --dump_clip ./clip_cache/WizardCoder-7B/int2-g128.pt 2. Get the Teacher Generation Databash

   vllm

   python generate_vllm.py --base_model --dataset_name code --out_path ./datasets/WizardCoder-7b/ --max_sample 3000 ```

   ```bash cd BitDistiller/data/generation

   bash generate.sh /root/WizardCoder-Python-7B/ code ../datasets/WizardCoder-7b/ 16 3000 3. Run KD-base QATbash

   Specify the pre-trained model path

   Specify the num_gpus and batch_size according to your GPU devices

   Specify the clipping cache path to the --clip

   cd train

   bash train.sh ../data/datasets/WizardCoder-7b/code_T0.7_N1024_S42_3000.json ./ckpts/WizardCoder-7b/int2-g128/ ./logs/WizardCoder-7b/int2-g128/ 2 ```

MetaMath

1. Get the Clipping result ```bash cd BitDistiller/quantization

   CUDA_VISIBLE_DEVICES=0 python autoclip.py --model_path --calib_dataset gsm8k --quant_type int --w_bit 2 --q_group_size 128 --run_clip --dump_clip ./clip_cache/MetaMath-7B/int2-g128.pt 2. Get the Teacher Generation Databash

   vllm

   python generate_vllm.py --base_model --dataset_name math --out_path ./datasets/MetaMath-7B/ --max_sample 3000 ```

   ```bash cd BitDistiller/data/generation

   bash generate.sh /root/MetaMath-7B-V1.0/ math ../datasets/MetaMath-7B/ 16 3000 3. Run KD-base QATbash

   Specify the pre-trained model path

   Specify the num_gpus and batch_size according to your GPU devices

   Specify the clipping cache path to the --clip

   cd train

   bash train.sh ../data/datasets/MetaMath-7B/math_T0.7_N1024_S42_3000.json ./ckpts/MetaMath-7b/int2-g128/ ./logs/MetaMath-7b/int2-g128/ 2 ```

3. Evaluation

Example Srcipts

LLaMA-2

• Test PPL on WikiText-2 ```bash cd test/general

python wiki_ppl.py --model ../../train/ckpts/hf-llama-2-7b/int2-g128/checkpoint-200/ --quant_type int --bits 2 --group_size 128 * Test MMLUbash CUDA_VISIBLE_DEVICES=0 python llm_eval.py --model ../../train/ckpts/hf-llama-2-7b/int2-g128/checkpoint-200/ --eval_tasks hendrycksTest-* --test_set --bits 2 --group_size 128 --quant_type int --num_fewshot 5 * Test Common-sense QA Tasksbash CUDA_VISIBLE_DEVICES=0 python llm_eval.py --model ../../train/ckpts/hf-llama-2-7b/int2-g128/checkpoint-200/ --eval_tasks arc_challenge,winogrande,hellaswag,piqa --test_set --bits 2 --group_size 128 --quant_type int --num_fewshot 0 ```

WizardCoder

• Install the environment according to the instructions of HumanEval,

• Example script: bash cd test/humaneval bash gen_preds.sh [checkpoint_path] ./preds/7b/int2-g128/

MetaMath

• Example script:

  bash cd test/gsm8k bash test.sh ../../train/ckpts/MetaMath-7b/int2-g128/ ./preds/7b/int2-g128/

4. Inference

Please see inference/

Reference

If you find BitDistiller useful or relevant to your research, please kindly cite our paper:

@misc{du2024bitdistiller,
      title={BitDistiller: Unleashing the Potential of Sub-4-Bit LLMs via Self-Distillation}, 
      author={Dayou Du and Yijia Zhang and Shijie Cao and Jiaqi Guo and Ting Cao and Xiaowen Chu and Ningyi Xu},
      year={2024},
      eprint={2402.10631},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}