把KV缓存内存调度化为几何问题，从理论界限到系统实践，打通LLM推理性能优化新路径。
arXiv:2606.22327v1 Announce Type: new Abstract: The explosive demand for interactive Large Language Model serving has highlighted the management of th…
针对大模型推理的尾部延迟问题，提出超越预测的尾感知调度方法，兼顾效率与公平。
arXiv:2606.18431v1 Announce Type: new Abstract: LLM serving exhibits extreme length variability, making size-based scheduling difficult in practice. R…
结合 AI 与排队论，提出新型调度设计，提升服务系统效率与公平性。
arXiv:2406.06855v3 Announce Type: replace-cross Abstract: To leverage prediction models to make optimal scheduling decisions in service systems, we mu…
面向Superchip的LLM推理优化，提出SLO感知的旋转调度与内存管理方案，已被MLSys '26接收。
arXiv:2601.20309v2 Announce Type: replace-cross Abstract: Large Language Model (LLM) serving faces a fundamental tension between stringent latency Ser…
从排队论视角揭秘LLM推理的吞吐最优调度算法，为系统优化提供数学根基
arXiv:2504.07347v3 Announce Type: replace-cross Abstract: As demand for Large Language Models (LLMs) and AI agents grows rapidly, optimizing systems f…