Instance-Aware Group Quantization for Vision Transformers

Post-training quantization works well for CNNs but breaks down with Vision Transformers due to highly variable activation distributions. IGQ-ViT solves this by dynamically grouping channels per input instance so each group shares similar statistics, then quantizing them with shared parameters. The method also extends to softmax attention and includes a group-allocation strategy under BOP constraints. Across classification, detection, and segmentation tasks, IGQ-ViT delivers state-of-the-art quantization results for ViTs at low bit-widths without costly retraining.