![[2602.12609] QuEPT: Quantized Elastic Precision Transformers with One-Shot Calibration for Multi-Bit Switching](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.12609] QuEPT: Quantized Elastic Precision Transformers with One-Shot Calibration for Multi-Bit Switching
Summary
The paper presents QuEPT, a novel quantization method for Transformers that enables efficient multi-bit switching with one-shot calibration, improving performance and adaptability in AI models.
Why It Matters
As AI models grow in complexity, optimizing their performance while reducing storage and computational costs is crucial. QuEPT addresses these challenges by enabling dynamic quantization, which can enhance model efficiency without sacrificing accuracy, making it relevant for developers and researchers in machine learning and AI.
Key Takeaways
- QuEPT allows for dynamic switching between different bit-widths in Transformer models.
- The method employs one-shot calibration to minimize optimization costs.
- Multi-Bit Token Merging (MB-ToMe) enhances feature robustness during quantization.
- The approach shows performance comparable to state-of-the-art quantization methods.
- QuEPT's design supports real-time adjustments, making it versatile for various applications.
Computer Science > Computer Vision and Pattern Recognition arXiv:2602.12609 (cs) [Submitted on 13 Feb 2026] Title:QuEPT: Quantized Elastic Precision Transformers with One-Shot Calibration for Multi-Bit Switching Authors:Ke Xu, Yixin Wang, Zhongcheng Li, Hao Cui, Jinshui Hu, Xingyi Zhang View a PDF of the paper titled QuEPT: Quantized Elastic Precision Transformers with One-Shot Calibration for Multi-Bit Switching, by Ke Xu and Yixin Wang and Zhongcheng Li and Hao Cui and Jinshui Hu and Xingyi Zhang View PDF HTML (experimental) Abstract:Elastic precision quantization enables multi-bit deployment via a single optimization pass, fitting diverse quantization this http URL, the high storage and optimization costs associated with the Transformer architecture, research on elastic quantization remains limited, particularly for large language this http URL paper proposes QuEPT, an efficient post-training scheme that reconstructs block-wise multi-bit errors with one-shot calibration on a small data slice. It can dynamically adapt to various predefined bit-widths by cascading different low-rank adapters, and supports real-time switching between uniform quantization and mixed precision quantization without repeated optimization. To enhance accuracy and robustness, we introduce Multi-Bit Token Merging (MB-ToMe) to dynamically fuse token features across different bit-widths, improving robustness during bit-width switching. Additionally, we propose Multi-Bit Cascaded Low-Rank adapters (M...
