ScalePredictor: Instance-aware Scale Learning for Accurate Quantization of Vision Transformers

ScalePredictor introduces a dynamic quantization framework for Vision Transformers (ViTs) aimed at improving post-training quantization (PTQ) efficiency. It leverages a correlation between shallow-layer activation distributions and optimal scales for deeper layers, employing a polynomial scale projection module for simultaneous quantization scale generation. This approach significantly enhances accuracy while minimizing computational overhead compared to existing static PTQ methods, making it particularly relevant for deploying ViTs on edge devices.