Reproduce “Kronecker Product Attention: Memory-Efficient KV Compression via Structured Factorization” 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
Bahareh Nikpour, Hossein Rajabzadeh, Maryam Dialameh, Vahid Partovi Nia, Parsa Omidi
KPA reduces language models’ key-value cache memory by storing matrix factors rather than full key and value matrices. Its factors extend Tensor Product Attention’s vector outer products at two structural levels. After about 50B training tokens, a 353M configuration matched TPA’s 6.4× compression and averaged 52.90 versus 51.41 across nine zero-shot tasks; these are benchmark and cache-storage results, not faster decoding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.