Beacon: Post-Training Quantization with Integrated Grid Selection

2025/08/27 by Zhang, Shihao, Saab, Rayan
#Artificial Intelligence (cs.AI) #FOS: Computer and information sciences #Machine Learning (cs.LG)

paper · doi:10.48550/arxiv.2508.20293

Abstract

Quantization is a widely used compression technique for reducing the memory and computation costs of large pre-trained models. A key challenge in per-channel post-training quantization (PTQ) is selecting appropriate scaling factors to replace weight values with values from a scaled integer grid. Existing methods typically fix the scale at the outset via heuristic tuning or grid search. We propose Beacon, a simple and effective algorithm that eliminates the need for such manual tuning. Beacon performs per-channel PTQ directly using an unscaled grid and automatically determines the optimal scaling factors by exploiting the geometry of scalar quantization. It does not rely on back-propagation or large calibration sets. Despite its simplicity and tuning-free nature, Beacon achieves competitive performance compared to state-of-the-art methods, making it a practical solution for efficient model deployment.

Citations

• Provable Post-Training Quantization: Theoretical Analysis of OPTQ and Qronos
• Qronos: Correcting the Past by Shaping the Future... in Post-Training Quantization
• Theoretical Guarantees for Low-Rank Compression of Deep Neural Networks
• MagR: Weight Magnitude Reduction for Enhancing Post-Training Quantization
• SpinQuant: LLM quantization with learned rotations
• LLMC: Benchmarking Large Language Model Quantization with a Versatile Compression Toolkit
• QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs
• COMQ: A Backpropagation-Free Algorithm for Post-Training Quantization
• Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs
• Training Transformers with 4-bit Integers
• AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration
• SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models
• GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers
• DeiT III: Revenge of the ViT
• Post-training Quantization for Neural Networks with Provable Guarantees
• ResNet strikes back: An improved training procedure in timm
• A Survey of Quantization Methods for Efficient Neural Network Inference
• Pruning and Quantization for Deep Neural Network Acceleration: A Survey
• Training data-efficient image transformers & distillation through attention
• A Greedy Algorithm for Quantizing Neural Networks
• Model Compression and Hardware Acceleration for Neural Networks: A Comprehensive Survey
• Quantization and Training of Neural Networks for Efficient\n Integer-Arithmetic-Only Inference

Related

• BoA: Attention-aware Post-training Quantization without Backpropagation
• PTQ1.61: Push the Real Limit of Extremely Low-Bit Post-Training Quantization Methods for Large Language Models
• AHCQ-SAM: Toward Accurate and Hardware-Compatible Post-Training Segment Anything Model Quantization
• Gradient-Based Post-Training Quantization: Challenging the Status Quo
• Solving Oscillation Problem in Post-Training Quantization Through a Theoretical Perspective