Breaking Modality Heterogeneity in Low-Bit Quantization for Large Vision-Language Models
Researchers have developed SplitQ, a new post-training quantization framework designed to improve the efficiency of large vision-language models (VLMs) on devices with limited resources. SplitQ addresses the accuracy degradation often seen in low-bit quantization by introducing a Modality-specific Outlier Channel Decoupling module to isolate modality-specific outliers and an Adaptive Cross-Modal Calibration module to correct remaining discrepancies. Experiments show SplitQ significantly outperforms existing methods across various quantization settings and datasets, preserving high performance even under challenging conditions. AI
IMPACT Enables more efficient deployment of advanced vision-language models on resource-constrained devices.