SeGO: Sensitivity-Aware Golden Optimization for Large-Scale VLM Quantization

摘要

The deployment of Vision-Language Models (VLMs) faces memory and computational bottlenecks because of the massive parameters and intensive computations. While Post-Training Quantization (PTQ) can reduce these costs, existing methods often overlook the heterogeneity of multimodal input when applied to VLMs, leading to quantization accuracy degradation. In this paper, we categorize the Transformer linear layers into Expansion Space (responsible for feature expansion) and Projection Space (responsible for feature aggregation) and reveal a cross-modal structural sensitivity asymmetry in VLMs: The Expansion Space is more sensitive to quantization accuracy than the Projection Space in both LLM and ViT encoders. Based on this, we propose SeGO, a unified structural sensitivity-aware sparse optimization framework. First, it constructs search interval for scaling factors using channel-wise statistics of LLM and ViT weights to exclude invalid solutions. Then it applies scaling protection only to the sensitive Expansion Space while directly quantizing the robust Projection Space. Moreover, SeGO uses a Golden-Section Solver to speedup the search of optimal scaling factors. Experiments show that SeGO achieves the balance among model parameter amount, quantization accuracy and scaling factors' search efficiency on InternVL2 and LLaVA series. Under the W3/W4A16 configuration, SeGO improves the accuracy of 7B-26B models by 3.8%-17.3% and enhances calibration efficiency by dozens of times.