Beyond Activation Alignment:The Alignment-Diversity Tradeoff in Task-Aware LLM Quantization

Fei Wang, Chao Xue, Taoran Liu +3

Mixed-precision quantization (MPQ) has become a key technique for deploying large language models under stringent memory and compute constraints. We first identify a phenomenon tha…