3 projects
locks-kv
LOCKS: Page-Local Compact Key Summaries for efficient long-context decoding -- an rki4 page-summary KV selector for vLLM (GQA + MLA/DeepSeek; optional DRAM value tier)
amass-kv
AMASS: decode-time sparse KV-cache attention for vLLM (quad page-score selector + DRAM value tier)
graphevo
Genetic Algorithm for Graph Optimization