About Me 个人简介
Hi there! I am currently a Ph.D. student in the School of Information at Renmin University of China, advised by Prof. Feng Zhang and Prof. Xiaoyong Du. My research interests lie in AI Infra and vector databases. I am broadly interested in building practical systems that improve throughput and cost efficiency under real-world resource constraints.
大家好!我目前是中国人民大学信息学院博士研究生,导师为张峰教授和杜小勇教授。我的研究方向是 AI Infra 和向量数据库。我对构建能够在实际资源约束下提升吞吐量和成本效率的实用系统有广泛兴趣。
News 近期动态
- Nov. 2026: Our native-H16 CuTe sparse-prefill backend for DeepSeek-V4 was merged into SGLang.
- Oct. 2026: Received the University Special Scholarship for Academic Excellence.
- Jul. 2026: Joined Alibaba Taotian Group as an AI Infra Engineer through the T-Star Lab Talent Program.
- Feb. 2026: TensorSlim was accepted at ICDE 2026.
- Sep. 2025: Received the China National Scholarship.
- Jun. 2025: HARMONY was accepted at SIGMOD 2026.
- Oct. 2024: Tribase was accepted at SIGMOD 2025.
- Sep. 2024: Received the China National Scholarship.
- Feb. 2024: Laconic was accepted at VLDB 2024.
- 2026 年 11 月:面向 DeepSeek-V4 的原生 H16 CuTe 稀疏预填充后端已合并至 SGLang。
- 2026 年 10 月:获得学习优秀学校特等奖学金。
- 2026 年 7 月:通过 T-Star Lab 人才计划加入阿里巴巴淘天集团,担任 AI Infra 工程师。
- 2026 年 2 月:TensorSlim 被 ICDE 2026 接收。
- 2025 年 9 月:获得研究生国家奖学金。
- 2025 年 6 月:HARMONY 被 SIGMOD 2026 接收。
- 2024 年 10 月:Tribase 被 SIGMOD 2025 接收。
- 2024 年 9 月:获得研究生国家奖学金。
- 2024 年 2 月:Laconic 被 VLDB 2024 接收。
Education 教育经历
- Renmin University of China, 2023.09 – Present
Ph.D. student in Computer Science - Renmin University of China, 2019.09 – 2023.06
B.Sc. in Computer Science
- 中国人民大学,2023.09 – 至今
计算机科学博士研究生 - 中国人民大学,2019.09 – 2023.06
计算机科学学士
Work Experience 工作经历
- Alibaba, Taotian Group, AI Infra Engineer, T-Star Lab Talent Program, 2026.07 – Present
- Ant Group, Research Intern, 2025.02 – 2026.06
Diffusion LLM AI Infra Department - Beijing HaiZhi XingTu Technology Co., Ltd, Research Intern, 2022.11 – 2025.01
Graph Computing Department - SenseTime, OCR Research Intern, 2021.11 – 2022.06
- 阿里巴巴淘天集团,AI Infra 工程师,T-Star Lab 人才计划,2026.07 – 至今
- 蚂蚁集团,研究实习生,2025.02 – 2026.06
扩散语言模型 AI 基础设施部门 - 北京海致星图科技有限公司,研究实习生,2022.11 – 2025.01
图计算部门 - 商汤科技,OCR 研究实习生,2021.11 – 2022.06
Publications 论文
- Qian Xu, Feng Zhang, Shijie Gao, Kun Chen, Jianhua Wang, Zheng Chen, Xiaoyong Du. TensorSlim: Improving GPU Tensor Query Processing for Resource-Constrained Environments. IEEE International Conference on Data Engineering (ICDE 2026). [Paper论文]
- Shijie Gao, Feng Zhang, Qian Xu, Yang Li, XueFeng Liu, Chao Jiang, Limin Xiao, Siqi Ma, Elisa Bertino, Xiaoyong Du. cuFHEDB: GPU-Accelerated Fully Homomorphic Encryption Database. IEEE International Conference on Data Engineering (ICDE 2026): 474–488. [Paper论文]
- Qian Xu, Feng Zhang, Chengxi Li, Lei Cao, Zheng Chen, Jidong Zhai, Xiaoyong Du. HARMONY: A Scalable Distributed Vector Database for High-Throughput Approximate Nearest Neighbor Search. ACM SIGMOD International Conference on Management of Data (SIGMOD 2026). [Paper论文]
- Yuxin Ma, Lun Du, Lanning Wei, Kun Chen, Qian Xu, Kangyu Wang, Guofeng Feng, Guoshan Lu, Lin Liu, Xiaojing Qi, Xinyuan Zhang, Zhen Tao, Haibo Feng, Ziyun Jiang, Ying Xu, Zenan Huang, Yihong Zhuang, Haokai Xu, Jiaqi Hu, Zhenzhong Lan, Junbo Zhao, Jianguo Li, Da Zheng. dInfer: An Efficient Inference Framework for Diffusion Language Models. Technical Report, 2025. [Paper论文]
- Qian Xu, Juan Yang, Feng Zhang, Junda Pan, Kang Chen, Youren Shen, Amelie Chi Zhou, Xiaoyong Du. Tribase: A Vector Data Query Engine for Reliable and Lossless Pruning Compression using Triangle Inequalities. ACM SIGMOD International Conference on Management of Data (SIGMOD 2025). [Paper论文]
- Thomas Bodner, Donghyun Sohn, Qian Xu. Reproducibility Report for ACM SIGMOD 2025 Paper: “Tribase: A Vector Data Query Engine for Reliable and Lossless Pruning Compression using Triangle Inequalities.” Reproducibility Reports of the ACM SIGMOD International Conference on Management of Data (SIGMOD ARI 2025): 59–64. [Paper论文]
- Qian Xu, Juan Yang, Feng Zhang, Zheng Chen, Jiawei Guan, Kang Chen, Ju Fan, Youren Shen, Ke Yang, Yu Zhang, Xiaoyong Du. Improving Graph Compression for Efficient Resource-Constrained Graph Analytics. 50th International Conference on Very Large Data Bases (VLDB 2024). [Paper论文]
Projects 项目
I contribute performance optimizations, correctness fixes, GPU kernels, and benchmark infrastructure to open-source AI systems.
- SGLang #36059Merged Built a native-H16 CuTe sparse-prefill backend for DeepSeek-V4 on NVIDIA SM90, avoiding padded-head computation and improving prefill efficiency.
- SGLang #48091Open Bounded SM90 MXFP4 preprocessing memory by chunking expert preparation, substantially reducing temporary VRAM use for large MoE checkpoints.
- SGLang #48073Open Fixed FP8 exclusion handling for Qwen4-Exp PLE tables so excluded parameters retain their intended BF16 precision.
我参与开源 AI 系统的性能优化、正确性修复、GPU 内核和基准测试基础设施建设。
- SGLang #36059已合并面向 NVIDIA SM90 为 DeepSeek-V4 实现原生 H16 CuTe 稀疏预填充后端,避免填充头的冗余计算并提升预填充效率。
- SGLang #48091进行中通过分块处理专家权重限制 SM90 MXFP4 预处理的内存峰值,显著降低大型 MoE 模型加载时的临时显存占用。
- SGLang #48073进行中修复 Qwen4-Exp PLE 表的 FP8 排除逻辑,使被排除参数保持预期的 BF16 精度。
dInfer is an open-source inference framework for diffusion language models. As a core contributor, I work on serving efficiency and KV-cache reuse.
- Vicinity Refresh. Designed a sliding-window policy that recomputes KV states near the active decoding block and reuses stable remote states, reducing redundant computation while preserving generation quality.
dInfer 是一个面向扩散语言模型的开源推理框架。作为核心贡献者,我主要参与推理服务效率和 KV 缓存复用优化。
- Vicinity Refresh。设计滑动窗口刷新策略,重新计算当前解码块附近的 KV 状态并复用远端稳定状态,在保持生成质量的同时减少冗余计算。
Honors & Awards 荣誉与奖励
- Young Science and Technology Scientists Sponsorship Program by CAST - Doctoral Student Special Plan, 2025
- China National Scholarship, 2025
- China National Scholarship, 2024
- University Special Scholarship for Academic Excellence, 2026
- 青年科技人才培育工程博士生专项计划(学院仅 3 人),2025 年
- 研究生国家奖学金,2025 年
- 研究生国家奖学金,2024 年
- 学习优秀学校特等奖学金,2026 年
Patents 专利
- ZL 2026 1 0305633.6: A tensor query execution optimization method, system, device, and medium for GPU resource-constrained scenarios. Inventors: Feng Zhang, Qian Xu, Xiaoyong Du.
- CN119759967A: A method and system for optimizing approximate nearest neighbor search of high-dimensional vectors based on compression technology. Inventors: Feng Zhang, Qian Xu, Xiaoyong Du.
- CN120086218A: A method, apparatus, medium, and device for constructing vector indexes based on sub-neighborhood networks. Inventors: Youren Shen, Qian Xu, Juan Yang, Yihua Liu, Zheng Wang.
- CN119169111A: A method for compressing and processing large-scale graph data in resource-constrained environments. Inventors: Youren Shen, Qian Xu, Juan Yang, Feng Zhang, Xiaoyong Du, Ke Yang.
- ZL 2026 1 0305633.6:一种面向 GPU 资源受限场景的张量查询执行优化方法、系统、设备和介质。发明人:张峰、许骞、杜小勇。
- CN119759967A:一种基于压缩技术优化高维向量近似最近邻搜索的方法和系统。发明人:张峰、许骞、杜小勇。
- CN120086218A:基于子邻近网络构建向量索引的方法、装置、介质和设备。发明人:沈游人、许骞、杨娟、刘艺华、王铮。
- CN119169111A:资源受限环境下对大规模图数据进行压缩和处理的方法。发明人:沈游人、许骞、杨娟、张峰、杜小勇、杨珂。