IGQ-ViT: Instance-Aware Group Quantization for Low-Bit Vision Transformers

IGQ-ViT introduces a dynamic channel-grouping strategy for quantizing Vision Transformers, improving accuracy and hardware efficiency with minimal latency trade-offs. Unlike previous methods that rely on static groups or fake quantization, it computes per-input statistical properties to assign channels more effectively, supports practical accelerator designs, and outperforms prior PTQ approaches on models like DETR.