SalQ-VLM: Fine-Grained Saliency-Guided Quantization for Vision-Language Models
This work proposes SalQ-VLM, an importance-aware PTQ framework that prioritizes salient tokens and suppresses redundant vision tokens during calibration, and derives a gradient-driven importance factor that captures token-level importance variance and is theoretically grounded in the relationship among loss perturbatio...
Yufei Xue, Yu-Shi Huang, Lu Zhu et al.
· 5 citations