量子位·2026-09-16 11:07·国内权威源把记忆交给CPU,大模型会变快中文摘要将KV缓存移至CPU并让GPU专攻Token生成,可大幅提升大模型运行效率。English SummaryOffloading memory to the CPU allows the GPU to focus on token generation, significantly accelerating large model performance.原文节选让GPU去忙生成Token阅读原文View in English