DAQ: Divide-and-Conquer Strategy Based Adaptive Low-Bit Quantization Method for Vision Transformer

Lü Qianru; Xu Jinwei; Jiang Jingfei; Li Dongsheng

doi:10.7544/issn1000-1239.202550145

Journal of Computer Research and Development > 2025 > Corrected proof > DOI: 10.7544/issn1000-1239.202550145 CSTR: 32373.14.issn1000-1239.202550145

Lü Qianru, Xu Jinwei, Jiang Jingfei, Li Dongsheng. DAQ: Divide-and-Conquer Strategy Based Adaptive Low-Bit Quantization Method for Vision Transformer[J]. Journal of Computer Research and Development. DOI: 10.7544/issn1000-1239.202550145

Citation:

PDF (1855 KB)

DAQ: Divide-and-Conquer Strategy Based Adaptive Low-Bit Quantization Method for Vision Transformer

National Key Laboratory of Parallel and Distributed Computing (National University of Defense Technology), Changsha 410073

Funds:

This work was supported by the National Natural Science Foundation of China (62025208, 62421002).

undefined

More Information

Author Bio:
Lü Qianru: born in 1993. PhD candidate. Her main research interests include computer architecture and artificial intelligence

Xu Jinwei: born in 1990. PhD, assistant research fellow. Member of CCF. His main research interests include artificial intelligence and reconfigurable computing

Jiang Jingfei: born in 1974. PhD, professor. Member of CCF. Her main research interests include reconfigurable computing, artificial intelligence, and computer architecture. （jingfeijiang@nudt.edu.cn）

Li Dongsheng: born in 1978. PhD, professor. Member of CCF. His main research interests include parallel computing, artificial intelligence, computer architecture.（dsli@nudt.edu.cn）
Received Date: February 28, 2025
Revised Date: April 09, 2025
Available Online: April 15, 2025

Graphical Abstract

Abstract

Abstract

Vision Transformers (ViTs) have demonstrated remarkable success in computer vision tasks, but their complex architecture and computational demands hinder deployment on edge devices. While post-training quantization (PTQ) is widely adopted for model compression, existing PTQ methods exhibit severe performance degradation in 4-bit ultra-low-bitwidth scenarios. This work systematically addresses two fundamental limitations: 1) spatial mismatch between quantization-sensitive layers (e.g., Softmax) and compute-intensive layers (e.g., linear projections), where quantizing Softmax causes 80% accuracy loss despite contributing merely 8% computational load; 2) non-Gaussian activation distributions with hidden Gaussian-like clustering properties (97% values less than three times z-score). We propose DAQ (divide-and-conquer and adaptive quantization), a hardware-friendly PTQ method. DAQ adpots z-score-driven dynamic partitioning algorithm to separate data normal-range and abnormal-range groups and quantizes the two groups with connected parameter. DAQ further explores hardware accelerated kernel such as tensor core to speed up quantization ViT models. Experimental results demonstrate that DAQ achieves a maximum improvement of 4.37% in ImageNet Top-1 accuracy under 4-bit quantization. In object detection tasks, its average error margin remains below 0.4% compared with the baseline and achieves a maximum improvement of 8.2%, even surpassing the full-precision model by 0.1% in specific cases, thereby realizing near-lossless low-bit-width quantization. Through hardware implementation optimization, DAQ achieves 43%~86% computational acceleration without significantly increasing computational overhead. This approach establishes a synergistic algorithm-hardware co-optimized quantization deployment paradigm for resource-constrained scenarios, effectively balancing model efficiency and precision retention.
- vision Transformer (ViT),
- post-training quantization (PTQ),
- outlier,
- low-bit quantization,
- z-score,
- uniform correlation quantization

FullText(HTML)

References (34)

References

[1]	Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. Advances in Neural Information Processing Systems, 2017, 30: 5998−6008
[2]	Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16×16 words: Transformers for image recognition at scale[J]. arXiv preprint, arXiv: 2010.11929, 2020
[3]	Touvron H, Cord M, Douze M, et al. Training data-efficient image transformers & distillation through attention[J]//arXiv: 2012.12877
[4]	Liu Ze, Lin Yutong, Cao Yue, et al. Swin Transformer: Hierarchical vision Transformer using shifted windows[C]//Proc of the IEEE/CVF Inte Conf on Computer Vision. Piscataway, NJ: IEEE, 2021: 10012−10022
[5]	He Kaiming, Zhang Xiangyu, Ren Shaoqing, et al. Deep residual learning for image recognition[C]//Proc of the IEEE Conf on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE, 2016: 770−778
[6]	Carion N, Massa F, Synnaeve G, et al. End-to-end object detection with transformers[C]//Proc of European Conference on Computer Vision. Berlin: Springer 2020: 213−229
[7]	Cheng B, Schwing A, Kirillov A. Per-pixel classification is not all you need for semantic segmentation[J]. Advances in Neural Information Processing Systems, 2021, 34: 17864−1787
[8]	Gündüç Y. Vit-GAN: Image-to-image translation with vision Transformes and conditional GANS[J]. arXiv preprint, arXiv: 2110.09305, 2021
[9]	Hatamizadeh A, Song J, Liu G, et al. Diffit: Diffusion vision transformers for image generation[C]//Proc of European Conf on Computer Vision. Berlin: Springer, 2024: 37−55
[10]	Lin Yang, Zhang Tianyu, Sun Peiqin, et al. Fq-ViT: Fully quantized vision Transformer without retraining[J]. arXiv preprint, arXiv: 2111.13824, 2021
[11]	Yuan Zhihang, Xue Chenhao, Chen Yiqi, et al. PTQ4ViT: Post-training quantization for vision transformers with twin uniform quantization[C]//Proc of European Conf on Computer Vision. Berlin: Springer, 2022: 191−207
[12]	Li Yanjing, Xu Sheng, Zhang Baochang, et al. Q-ViT: Accurate and fully quantized low-bit vision Transformer[J]. Advances in Neural Information Processing Systems, 2022, 35: 34451−34463
[13]	Zhong Yunshan, Huang You, Hu Jiawei, et al. Towards accurate post-training quantization of vision transformers via error reduction[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025, 47(4): 2676−2692
[14]	Li Zhikai, Xiao Junrui, Yang Lianwei, et al. RepQ-ViT: Scale reparameterization for post-training quantization of vision transformers[C]//Proc of the IEEE/CVF Int Conf on Computer Vision. Piscataway, NJ: IEEE, 2023: 17227−17236
[15]	Li Zhikai, Liu Xuewen, Zhang Jing, et al. RepQuant: Towards accurate post-training quantization of large Transformer models via scale reparameterization[J]. arXiv preprint, arXiv: 2402.05628, 2024
[16]	Du Dayou, Gong Gu, Chu Xiaowen. Model quantization and hardware acceleration for vision transformers: A comprehensive survey[J]. arXiv preprint, arXiv: 2405. 00314, 2024
[17]	Tai Yushan, Lin Mingguang, Wu A. TSPTQ-ViT: Two-scaled post-training quantization for vision Transformer[C]//Proc of 2023 IEEE Int Conf on Acoustics, Speech and Signal Processing (ICASSP). Piscataway, NJ: IEEE, 2023: 1−5
[18]	Chen C F R, Fan Q, Panda R. Crossvit: Cross-attention multi-scale vision transformer for image classification[C]//Proc of the IEEE/CVF Int Conf on computer vision. Piscataway, NJ: IEEE, 2021: 357−366
[19]	Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84−90
[20]	Zhu Xizhou, Su Weijie, Lu Lewei, et al. Deformable DETR: Deformable transformers for end-to-end object detection[J]. arXiv preprint, arXiv: 2010.04159, 2020
[21]	Strudel R, Garcia R, Laptev I, et al. Segmenter: Transformer for semantic segmentation[C]//Proc of the IEEE/CVF Int Conf on Computer Vision. Piscataway, NJ: IEEE, 2021: 7262−7272
[22]	Zheng Sixiao, Lu Jiachen, Zhao Hengshuang, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers[C]//Proc of the IEEE/CVF Conf on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE, 2021: 6881−6890
[23]	Liu Yixin, Zhang Kai, Li Yuan, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models[J]. arXiv preprint, arXiv: 2402.17177, 2024
[24]	Dettmers T, Lewis M, Belkada Y, et al. GPT3. Int8 (): 8-bit matrix multiplication for transformers at scale[J]. Advances in Neural Information Processing Systems, 2022, 35: 30318−30332
[25]	Touvron H, Lavril T, Izacard G, et al. Llama: Open and efficient foundation language models[J]. arXiv preprint, arXiv: 2302.13971, 2023
[26]	Chan T F, Golub G H, LeVeque R J. Algorithms for computing the sample variance: Analysis and recommendations[J]. The American Statistician, 1983, 37(3): 242−247 doi: 10.1080/00031305.1983.10483115
[27]	Nvidia. NVIDIA GPU Ampere Architecture Whitepaper, https://www.nvidia.com/content/PDF/nvidia-ampere-ga-102-gpu-architecture-whitepaper-v2.pdf
[28]	Deng Jia, Dong Wei, Socher R, et al. ImageNet: A large-scale hierarchical image database[C]//Proc of 2009 IEEE Conf on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE, 2009: 248−255
[29]	Lin Tsung-Yi, Maire M, Belongie S, et al. Microsoft COCO: Common objects in context[C]//Proc of the 13th European Conf on Computer Vision (ECCV 2014). Berlin: Springer, 2014: 740−755
[30]	Chen Kai, Wang Jiaqi, Pang Jiangmiao, et al. MMDetection: Open mmlab detection toolbox and benchmark[J]. arXiv preprint, arXiv: 1906.07155, 2019
[31]	Frantar E, Ashkboos S, Hoefler T, et al. GPTQ: Accurate post-training quantization for generative pre-trained transformers[J]. arXiv preprint, arXiv: 2210.17323, 2022
[32]	Nagel M, Amjad R A, Van Baalen M, et al. Up or down? adaptive rounding for post-training quantization[C]//Proc of Int Conf on Machine Learning. New York: ACM, 2020: 7197−7206
[33]	Peebles W, Xie S. Scalable diffusion models with transformers[C]//Proc of the IEEE/CVF Int Conf on Computer Vision. Piscataway, NJ: IEEE, 2023: 4195−4205
[34]	Ho J, Jain A, Abbeel P. Denoising diffusion probabilistic models[J]. Advances in Neural Information Processing Systems, 2020, 33: 6840−6851

[1]	Qian Zhongsheng, Huang Heng, Zhu Hui, Liu Jinping. Multi-Perspective Graph Contrastive Learning Recommendation Method with Layer Attention Mechanism[J]. Journal of Computer Research and Development, 2025, 62(1): 160-178. DOI: 10.7544/issn1000-1239.202330804
[2]	Li Minghui, Jiang Peipei, Wang Qian, Shen Chao, Li Qi. Adversarial Attacks and Defenses for Deep Learning Models[J]. Journal of Computer Research and Development, 2021, 58(5): 909-926. DOI: 10.7544/issn1000-1239.2021.20200920
[3]	Du Peng, Ding Shifei. A DGA Domain Name Detection Method Based on Deep Learning Models with Mixed Word Embedding[J]. Journal of Computer Research and Development, 2020, 57(2): 433-446. DOI: 10.7544/issn1000-1239.2020.20190160
[4]	Wang Ruiqin, Wu Zongda, Jiang Yunliang, Lou Jungang. An Integrated Recommendation Model Based on Two-stage Deep Learning[J]. Journal of Computer Research and Development, 2019, 56(8): 1661-1669. DOI: 10.7544/issn1000-1239.2019.20190178
[5]	Li Hongshun, Yu Hua, Gong Xiujun. A Deep Learning Model for Predicting RNA-Binding Proteins Only from Primary Sequences[J]. Journal of Computer Research and Development, 2018, 55(1): 93-101. DOI: 10.7544/issn1000-1239.2018.20160508
[6]	Gong Xiaoli, Yu Haiyang, Sun Chengjun, Li Tao, Zhang Jin, Ma Jie. XOS: A QoE Oriented Energy Efficient Heterogeneous Multi-Processor Schedule Mechanism[J]. Journal of Computer Research and Development, 2016, 53(7): 1467-1477. DOI: 10.7544/issn1000-1239.2016.20160113
[7]	Wu Lei, Zhang Wensheng, Wang Jue. Hidden Topic Variable Graphical Model Based on Deep Learning Framework[J]. Journal of Computer Research and Development, 2015, 52(1): 191-199. DOI: 10.7544/issn1000-1239.2015.20131113
[8]	Zhu Xia, Song Aibo, Dong Fang, Luo Junzhou. A Collaborative Filtering Recommendation Mechanism for Cloud Computing[J]. Journal of Computer Research and Development, 2014, 51(10): 2255-2269. DOI: 10.7544/issn1000-1239.2014.20130056
[9]	Liu Yuheng, Wu Jing, Chen Zhenyong, Xiong Zhang. A Dynamic Sleep Scheduling Mechanism for Localization in Mobile Sensor Networks[J]. Journal of Computer Research and Development, 2008, 45(8): 1330-1337.
[10]	Zhu Jinghua, Li Jianzhong, Liu Yong, and Gao Hong. Data-Driven Sleeping Scheduling Mechanism in Sensor Networks[J]. Journal of Computer Research and Development, 2008, 45(1): 172-179.