Reproduce “Merge-Friendly Post-Training Quantization for Multi-Target Domain Adaptation” on OpenResearch

About this paper

Juncheol Shin, Minsang Seok, Seonggon Kim, Eunhyeok Park

POSTECH researchers develop HDRQ (Hessian and Distance Regularizing Quantization), the first post-training quantization method designed specifically for model merging in multi-target domain adaptation scenarios, addressing how quantization noise disrupts weight alignment between domain-adapted models by introducing noise-based Hessian regularization to flatten loss surfaces and distance regularization to maintain proximity to source weights, achieving substantial improvements in merged model performance at lower bit-widths (4.21 mIoU gain over QDrop at W4A4 in semantic segmentation and 3.42% improvement in Office-Home classification) through theoretical analysis extending the error barrier concept to quantized model spaces.

Key insights

• Theoretical extension of error barrier analysis to quantized weights reveals that minimizing quantization impact on merging requires reducing Hessian magnitude (sharper landscapes amplify noise effects) and controlling weight divergence between models
• Flatter loss surfaces achieved through noise-based regularization are crucial for stable model merging, as visualized through loss landscape analysis showing HDRQ converges to smoother minima compared to existing methods

Summary by alphaXiv.