Reproduce “TenureKV: Promotion-Based Precision Allocation for Long-Context Key-Value Caches” 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
0xb0ba
TenureKV addresses long-context inference memory pressure by periodically reviewing groups of cached tokens and retaining full-precision copies in host memory so pages can be restored if later needed. In reported LongBench results on Llama-3.1-8B-Instruct, it scored 48.55, versus PyramidKV’s 47.45 at equal 17.5% GPU memory. The preprint says its numbers were obtained without a GPU, so they are unverified.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.