BoA

This repository contains the code for the ICML 2025 paper BoA: Attention-Aware Post-Training Quantization without Backpropagation.

The current release includes the following features:

• Implementation of the proposed BoA: boa.py
• Quantization of OPT, Llama, Llama2, Llama3, Qwen2.5, Qwen3 models: main.py
• Evaluating the perplexity and 0-shot accuracy (8 tasks) of quantized models

Dependencies

• see requirements.txt

BoA options

• block_v: whether to apply block-wise objective for the value projection layer. In memory-limited cases, we can significantly reduce memory by de-activating this option, but at the expense of a slight performance degradation.
• act_order_col: whether to re-order columns before the quantization based on the column-wise Hessian $\mathbf{H}_{col}$ (GPTQ heuristic)
• act_order_row: whether to re-order rows before the quantization based on the row-wise Hessian $\mathbf{H}_{row}$
• qparam_comput: how to select quantization grids. Grids can be determined with a naive MinMax or to minimize the weight perturbation (MMSE) or the layer-wise reconstruction error (Hessian)

For more details on other arguments, please refer to process_args.py.

Experimental Results

• Setup
  • NVIDIA H100 GPU has been used.
  • block_v option has been activated.
  • qparam_comput option has been set to Hessian.
  • Test all cases for act_order_row and act_order_col and report the best results with respect to Wiki2 PPL.

Results on Qwen2.5 Models

• INT2 weight-only quantization

  Size	act_order_row	act_order_col	Wiki2 ($\downarrow$)	C4-new ($\downarrow$)	0-shot ($\uparrow$)
  0.5B	O	O	144.7	455.8	32.56
  1.5B	O	O	58.09	235.7	36.93
  3B	X	O	26.55	90.77	43.28
  7B	X	O	23.14	103.4	43.79
  14B	O	O	12.05	37.64	57.4

• INT3 weight-only quantization

  Size	act_order_row	act_order_col	Wiki2 ($\downarrow$)	C4-new ($\downarrow$)	0-shot ($\uparrow$)
  0.5B	O	O	20.12	40.14	45.68
  1.5B	X	O	12.03	23.52	56.90
  3B	O	O	9.541	17.32	60.29
  7B	O	O	9.054	18.19	66.12
  14B	O	O	6.492	12.20	71.80

• Quantization processing time

  Size	Time (min)
  0.5B	6.145
  1.5B	22.81
  3B	39.59
  7B	63.14
  14B	150.9

Results on Qwen3 Models

• INT2 weight-only quantization

  Size	act_order_row	act_order_col	Wiki2 ($\downarrow$)	C4-new ($\downarrow$)	0-shot ($\uparrow$)
  4B	X	O	78.57	302.8	35.13
  8B	X	O	32.53	96.43	38.47
  14B	X	O	24.29	75.76	42.41

• INT3 weight-only quantization

  Size	act_order_row	act_order_col	Wiki2 ($\downarrow$)	C4-new ($\downarrow$)	0-shot ($\uparrow$)
  4B	X	O	28.19	48.24	45.25
  8B	X	O	15.62	29.60	53.69
  14B	X	O	10.62	18.61	66.69

• Quantization processing time

  Size	Time (min)
  4B	43.10
  8B	83.77
  14B	132.6

License

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC).

Citation

If you find this work is useful for your research, please cite our paper:

@inproceedings{kimboa,
  title={BoA: Attention-aware Post-training Quantization without Backpropagation},
  author={Kim, Junhan and Kim, Ho-young and Cho, Eulrang and Lee, Chungman and Kim, Joonyoung and Jeon, Yongkweon},
  booktitle={Forty-second International Conference on Machine Learning}
}