Reproduce “LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers” 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
Tianle Zhou, Yang Zhang, Tianyu Wu, Zhuohan Wang, William Nixon, Juncheng Yang
LoopKV reduces key–value cache reads in looped transformers by profiling each layer-loop cell and choosing INT4, INT2, or sharing, then using each loop’s attention scores to select values for full precision at read time. On Ouro-1.4B, it delivers 1.7–2.2× KIVI-4 goodput, with no public-suite average significantly below BF16; variable tracking declines at 32K.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.