vLLM (original release)
paperPagedAttention and the original vLLM serving engine (SOSP 2023, Kwon/Li/Zhuang/Sheng/Zheng/Stoica) — the paper that defined modern LLM serving. Pre-window legacy anchor, pinned per our attribution rule: the living vLLM project (86K★, PyTorch Foundation, 2,000+ contributors, Inferact commercialization) is not university output; Berkeley keeps this original dated artifact. SGLang (2312.07104) has the same shape: original Berkeley/Stanford artifact, community-project evolution.