Reproduce “Virtual Head Attention” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xibo Ding, Guoxia Wang, Jinle Zeng, Jiabin Yang, Sirui Tao, Chao Yang, Dianhai Yu
VHA changes attention parameterization during training so a compressed model can recover attention flexibility without extra serving operations. It uses learned linear query transformations and head-output mixing, which can be folded into standard GQA-2. In 1.7B models, its 8-2-2 configuration scores 54.50 across eight zero-shot benchmarks, versus 52.38 for ordinary 8-2-2 GQA, under the reported training and evaluation setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.