arrow
Return

Token-based dynamic bit-width assignment for ViT quantization

delete2025-08-16
delete0
PRE
AI
D
Dohyung Kim
J
Jaehyeon Moon
J
Junghyup Lee
G
Geon Lee
J
Jeimin Jeon
B
Bumsub Ham *
DOI:10.1016/j.patcog.2025.112269delete
deleteOriginal
deleteOriginal request for help
deleteShare
deleteSave
Abstract

Abstract

En 中文
• We introduce a novel token-based dynamic bit-allocation method for ViT quantization that assigns bit-widths of individual tokens dynamically, including the class token, for each input instance. • We present a shifting quantizer, specially designed for post-GELU activations, that considers the asymmetric distribution across positive and negative values. • We demonstrate that our approach boosts the performance of the token-wise quantization scheme significantly, and show that it achieves a good compromise between efficiency and efficacy.
Keywords:
token-based dynamic bit-allocation
ViT quantization
shifting quantizer
post-GELU activations
efficiency and efficacy

Journal

Pattern Recognition cover
Pattern Recognition
IF:
7.6
Papers:
1.3W
Citations:
4.5W

Organization

No organization information available
Cited Papers

Cited Papers

Loss aware post-training quantization
err2021-10-01
err54
errOAAI
errNahshan, Yury; Chmiel, Brian; Baskin, Chaim; Zheltonozhskii, Evgenii; Banner, Ron; Bronstein, Alex M.; Mendelson, Avi
errShare
errSave
Mixed-precision quantized neural networks with progressively decreasing bitwidth
err2021-03-01
err27
PREAI
errChu, Tianshu; Luo, Qin; Yang, Jie; Huang, Xiaolin
errShare
errSave
Data-free quantization via mixed-precision compensation without fine-tuning
err2023-11-01
err13
errOAAI
errChen, Jun; Bai, Shipeng; Huang, Tianxin; Wang, Mengmeng; Tian, Guanzhong; Liu, Yong
errShare
errSave