On-Policy Morphing into Hybrid Attention Models
Disen Lan, Xuanda Wang, Jianbin Zheng, Huafeng Kuang, Wei Tang, Yu Cheng
原文摘要
Hybrid attention models combine full and linear attention to achieve a better quality--efficiency trade-off. Converting pretrained Transformers into such models requires identifying which layers should retain full attention under a prescribed budget. We identify two off-policy mismatches in FlashMorph's layer selection: (i) continuous branch interpolation and teacher-forced hidden states differ from discrete hybrid computation; and (ii) frozen model weights fail to account for parameter adaptation during subsequent recovery. We introduce FlashMorph2, an on-policy layer-selection method that (I) evaluates sampled discrete hybrid architectures end to end using their own hidden states, and (II) jointly optimizes layer-selection scores and model weights under the recovery objective to reflect post-adaptation performance. Experiments across Qwen3 model scales and linear-attention variants show that FlashMorph2 consistently outperforms existing layer-selection methods, improving average scores over FlashMorph by $+2.42$ points on RULER, $+1.08$ on LongBench, and $+0.76$ on recall-intensive tasks, while completing layer selection in only $0.72$--$3.57$ GPU hours across 0.6B--8B models, demonstrating high efficiency and scalability.
cs.CL