Qinyu Xu

Ph.D. Student
Courant Institute of Mathematical Sciences
New York University

EmailGoogle ScholarGitHubLinkedIn
Qinyu Xu

About

I am a Ph.D. student in Computer Science at the Courant Institute of Mathematical Sciences, New York University, working with Anirudh Sivaraman. Before NYU, I received my B.E. in Computer Science and Engineering from Tsinghua University. I work on distributed systems and formal verification, and how both can make machine learning systems more reliable.

Research Interests

  • Distributed systems
  • Formal verification
  • Machine learning systems

Research

vCheck: Verifiable Checkpointing for LLM Training
New York University, 2025 – Present. Advisor: Anirudh Sivaraman.

  • Designed a two-layer checkpointing architecture that separates a general, application-oblivious consistency argument from application-specific state saving.
  • Completed a verified implementation of the application-oblivious layer: wrote the generic library in Go, translated it into Rocq via Goose, and proved in Grove that it satisfies the system’s state reachability theorem.
  • Adapted the Chandy–Lamport snapshot algorithm to capture globally consistent checkpoints asynchronously, without stalling training or requiring programmable-switch support.

Tactic: Sparse Attention for Long-Context LLM Inference
University of Washington, 2024 – 2025. Supervisor: Baris Kasikci.

  • Proposed and developed a dynamic token-selection method for KV cache optimization in long-context LLM inference.
  • Designed a three-phase approach based on spherical k-means clustering, token ordering, and attention-score distribution modeling.
  • Developed a custom kernel to accelerate k-means clustering performance.

NanoFlow: High-Throughput LLM Serving
University of Washington, 2024 – 2025. Supervisor: Baris Kasikci.

  • Profiled GPU utilization and operator-level performance for the serving pipeline.

Open-Structure Table Extraction
Microsoft Research Asia, 2023 – 2024. Supervisor: Haoyu Dong.

  • Built a data pipeline transforming large open-domain corpora into high-quality annotated tables.
  • Designed an evaluation framework for benchmarking models on open-structure table detection.

Education

New York University — New York, NY
Ph.D. in Computer Science, 2025 – Present

Tsinghua University — Beijing, China
B.E. in Computer Science and Engineering, 2021 – 2025

Publications and Posters

vCheck: Building Efficient and Verifiable Checkpointing for LLM Training
Qinyu Xu, Jinkun Geng, Joseph Tassarotti, Anirudh Sivaraman
SOSP 2026 Poster, Prague, Czech Republic

Tactic: Adaptive Sparse Attention with Clustering and Distribution Fitting for Long-Context LLMs
Kan Zhu, Tian Tang, Qinyu Xu, Zhan Jin, Yile Gu, Zhichen Zeng, Rohan Kadekodi, Liangyu Zhao, Ang Li, Arvind Krishnamurthy, Baris Kasikci
ICLR 2026, Rio de Janeiro, Brazil
Paper

NanoFlow: Towards Optimal Large Language Model Serving Throughput
Kan Zhu, Yufei Gao, Yilong Zhao, Liangyu Zhao, Gefei Zuo, Yile Gu, Dedong Xie, Tian Tang, Qinyu Xu, Zihao Ye, Keisuke Kamahori, Chien-Yu Lin, Ziren Wang, Stephanie Wang, Arvind Krishnamurthy, Baris Kasikci
OSDI 2025, Boston, MA, USA
Paper

OpenTE: Open-Structure Table Extraction from Text
Haoyu Dong, Mengkang Hu, Qinyu Xu, Haochen Wang, Yue Hu
ICASSP 2024, Seoul, Korea