FBI-LLM: Scaling Up Fully Binarized LLMs from Scratch via Autoregressive Distillation
Liqun Ma, Mingjie Sun, Zhiqiang Shen
Mohamed bin Zayed University of Artificial Intelligence.
Carnegie Mellon University.
This work presents a Fully BInarized Large Language Model (FBI-LLM), demonstrating for the first time how to train a large-scale binary language model (not the ternary LLM like BitNet b1.58 from scratch to match the performance of its full-precision counterparts (e.g., FP16 or BF16) in transformer-based LLMs. It achieves this by employing an autoregressive distillation (AD) loss with maintaining equivalent model dimensions (130M, 1.3B, 7B) and training data volume as regular LLM pretraining, while delivering competitive results in terms of perplexity and task-specific effectiveness. Intriguingly, by analyzing the training trajectory, we find that the pretrained weight is not necessary for training binarized LLMs from scratch. This research encourages a new computational framework and may facilitate the future design of specialized hardware tailored for fully 1-bit LLMs. We make all models, code, and training dataset fully accessible and transparent to support further research.
torch==2.1.1
transformers==4.40.0
tokenizers==0.19.1
datasets==2.19.0
lightning==2.1.3
flash-attn==2.5.0
fastchat==0.1.0
lm-eval==0.4.2
pytz
wandb
fire
You can use sbatch
to submit slurm job to train FBI-LLM:
sbatch sbatch_train.sh
Please modify the corresponding parameters in sbatch_train.sh
according to the configuration of your cluster.
Pretrained moodels are released on HuggingFace (https://huggingface.co/LiqunMa/): FBI-LLM-130M
, FBI-LLM-1.3B
, FBI-LLM-7B
.
The structural parameters of these models are as follows:
FBI-LLM 130M | FBI-LLM 1.3B | FBI-LLM 7B | |
---|---|---|---|
# layers | 12 | 24 | 32 |
hidden size | 768 | 2048 | 4096 |
intermediate size | 2048 | 5632 | 11008 |
# attention heads | 12 | 32 | 32 |
We use lm-evaluation-harness to evaluate FBI-LLMs.
lm-evaluation-harness
(pip install lm-eval==0.4.2
)load_ckpts()
from eval.py
to load the model. If you want to directly evaluate our open-sourced model, please use the following command:
sbatch sbatch_eval.sh
If you find our work useful and helpful to your research, please consider citing this paper:
@misc{ma2024fbillmscalingfullybinarized,
title={FBI-LLM: Scaling Up Fully Binarized LLMs from Scratch via Autoregressive Distillation},
author={Liqun Ma and Mingjie Sun and Zhiqiang Shen},
year={2024},
eprint={2407.07093},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2407.07093},
}