TP-ViT: truncated uniform-log2 quantizer and progressive bit-decline reconstruction for vision Transformer quantization

• Introduces a truncated uniform-log2 quantizer that effectively handles outliers in post-Softmax activations, significantly reducing quantization errors. • Proposes a progressive bit-decline optimization strategy that gradually reduces bit precision while preserving model performance under extreme 3-bit quantization. • Achieves a notable 6.18 percentage points improvement in top-1 accuracy for ViT-small under 3-bit quantization, outperforming state-of-the-art PTQ methods. • Demonstrates robust performance across image classification, object detection, and instance segmentation, enabling efficient deployment of ViTs on edge hardware.