¹The University of Hong Kong ² Beihang University ³ETH Zürich

Pretrained large language models (LLMs) exhibit exceptional general language processing capabilities but come with significant demands on memory and computational resources. As a powerful compression technology, binarization can extremely reduce model weights to a mere 1 bit, lowering the expensive computation and memory requirements. However, existing quantization techniques fall short of maintaining LLM performance under ultra-low bit-widths. In response to this challenge, we present BiLLM, a groundbreaking 1-bit post-training quantization scheme tailored for pretrained LLMs. Based on the weight distribution of LLMs, BiLLM first identifies and structurally selects salient weights, and minimizes the compression loss through an effective binary residual approximation strategy. Moreover, considering the bell-shaped distribution of the non-salient weights, we propose an optimal splitting search to group and binarize them accurately. BiLLM achieving for the first time high-accuracy inference (e.g. 8.41 perplexity on LLaMA2-70B) with only 1.08-bit weights across various LLMs families and evaluation metrics, outperforms SOTA quantization methods of LLM by significant margins. Moreover, BiLLM enables the binarization process of the LLM with 7 billion weights within 0.5 hours on a single GPU, demonstrating satisfactory time efficiency.

• [2024/2] BiLLM source code is open now!

• torch: tested on v2.0.1+cu117
• transformers: tested on v4.35.0 (the LLaMa integration currently requires a main install from source and sentencepiece)
• datasets: tested on v2.14.6
• huggingface-hub: tested on v0.16.4

All binarization process and experiments were run on a single 80GB NVIDIA A100. However, all the process can also be conducted on a single 24GB NVIDIA 3090 Ti when the model's parameter is under 70B.

python3 run.py facebook/opt-6.7b c4 braq --blocksize 128 --salient_metric hessian

python3 run.py meta-llama/Llama-2-7b-hf c4 braq --blocksize 128 --salient_metric hessian

or

python3 run.py huggyllama/llama-7b c4 braq --blocksize 128 --salient_metric hessian

python3 run.py lmsys/vicuna-7b-v1.5 c4 braq --blocksize 128 --salient_metric hessian

• BiLLM achieve superior perplexity performance on Wikitext2 datasets within only an average of 1.11 bit-width weights OPT families.

• BiLLM achieve superior perplexity performance on Wikitext2 datasets within only an average of 1.09 bit-width weights LLaMA families and 1.08 bit-width weights LLaMA2 families.

• We also evaluated the performance of BiLLM on PTB and C4 datasets.

• We further evaluated BiLLM on 7 zero-shot dataset to give extensive insight on binarization LLMs

  

• BiLLM achieve superior perplexity performance on Wikitext2 datasets within only an average of 1.10 bit-width weights Vicuna families (instruction fine-tune models).

If you find BiLLM is useful and helpful to your work, please kindly cite this paper:

@misc{huang2024billm,
      title={BiLLM: Pushing the Limit of Post-Training Quantization for LLMs}, 
      author={Wei Huang and Yangdong Liu and Haotong Qin and Ying Li and Shiming Zhang and Xianglong Liu and Michele Magno and Xiaojuan Qi},
      year={2024},
      eprint={2402.04291},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}