What Makes Vision Transformers Hard to Quantize?

This article reviews key approaches to neural network quantization, comparing QAT and PTQ techniques while highlighting why transformer architectures—especially Vision Transformers—pose unique challenges for low-bit conversion. It summarizes prior solutions across CNNs and ViTs, including channel-wise, layer-wise, and group-based quantizers, and explains how recent research attempts to handle extreme inter-channel scale variations. The piece concludes by contrasting static grouping methods with a newer dynamic, per-instance group quantization strategy designed to preserve accuracy in ViTs without additional training parameters.