QuantGen: Parameter Generation for Controllable Model Quantization

Mengxuan Wu, Xinhao Zhong, Penghao Wang, Ziheng Qin, Jimeng Zhang, Wangbo Zhao, Mingjia Shi, Zekai Li, Hao Ye, Panpan Zhang, Shuo Wang, Yang You, Kai Wang, Dawei Yang

Keywords: parameter generation, quantization

TL;DR: We extend the capabilities of parameter generation by applying it to the domain of model quantization

Abstract: Parameter generation has recently emerged as a novel and effective paradigm in the pursuit of efficient AI, offering a fundamentally different perspective from conventional deep learning by directly synthesizing high-quality model parameters. Despite its promise, existing parameter generation methods are typically constrained to producing parameters aligned with the task objectives present in their training data. This limitation significantly restricts their applicability and practical utility across diverse real-world scenarios. In this work, we introduce a dedicated parameter generation method specifically designed for model quantization—a critical step in deploying deep learning models on resource-constrained devices. We propose the first recurrent-based generator capable of directly producing quantized model parameters that retain performance comparable to their full-precision counterparts, without requiring any additional data or retraining. Furthermore, our framework supports controllable quantization, enabling the generation of parameters that satisfy varying precision and deployment requirements. Extensive experiments across multiple datasets and model architectures demonstrate that our method achieves strong generalization and robustness under a wide range of quantization settings. These findings underscore the potential of parameter generation as a powerful and flexible tool for efficient model compression, training, and deployment.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 17751