Zihan Zhang | 章子涵

I'm a fourth-year undergraduate in the ACM Honors Class at Zhiyuan College, Shanghai Jiao Tong University (B.Eng. in CS, expected 2027). GPA: 3.90/4.3 (Rank 11/27).

My research focuses on efficient machine learning systems, particularly low-bit quantization and numerics for large language models (LLMs) and efficient LLM serving via speculative decoding.

Currently a visiting student at Z-Lab (UC San Diego, advised by Prof. Zhijian Liu). Previously a research intern at EPCC Lab (SJTU, advised by Prof. Jingwen Leng).

Check my CV or email me!

Email  /  CV  /  GitHub  /  X


profile photo
News
  • [Nov 2026] Emulating Tensor-Core Numerics for Reproducible Low-Bit LLM Inference accepted to HPCA 2027!
  • [Oct 2026] DFlash2 submitted to MLSys 2027. arXiv coming soon.
  • [Feb 2026] ParoQuant accepted to ICLR 2026! Work done as a visiting student at Z-Lab (UC San Diego).
  • [Jan 2026] M2XFP accepted to ASPLOS '26 (EPCC Lab project).
  • [Oct 2025] Joined Z-Lab (UC San Diego) as a visiting student, working on efficient LLM quantization (ParoQuant).
Open-Source Contributions
  • DFlash2 upstreamed to vLLM, SGLang and llama.cpp (all merged)
    vLLM #52816 (bugfixes #54373, #54374)  /  SGLang #35371  /  llama.cpp #27342
  • Follow-up: shared-KV drafts for SGLang (#47982, under review).
Publications
  • DFlash2: Efficient Parallel Drafting for LLM Serving
    Jian Chen, Zihan Zhang, Zhijian Liu
    Under review at MLSys 2027   [arXiv coming soon]
  • Emulating Tensor-Core Numerics for Reproducible Low-Bit LLM Inference
    Weiming Hu, Zihan Zhang, Zihan Liu, ... Minyi Guo, Jingwen Leng
    HPCA 2027
  • M2XFP: A Metadata-Augmented Microscaling Data Format for Efficient Low-bit Quantization
    Weiming Hu, Zihan Zhang, Haoyan Zhang, ... Jingwen Leng
    ASPLOS 2026   [arXiv]   [code]
  • ParoQuant: Pairwise Rotation Quantization for Efficient Reasoning LLM Inference
    Yesheng Liang, Haisheng Chen, Zihan Zhang, Song Han, Zhijian Liu
    ICLR 2026   [arXiv]   [project]   [code]

Research & Education Experience
Z-Lab, UC San Diego
Oct 2025 – Present
Visiting Student
Advisor: Prof. Zhijian Liu
Speculative decoding for LLM serving (DFlash2, under review at MLSys 2027); pairwise rotation quantization for reasoning LLMs (ParoQuant, ICLR 2026).
Emerging Parallel Computer Center (EPCC Lab), SJTU
Jun 2025 – May 2026
Research Intern
Advisor: Prof. Jingwen Leng
Low-bit data formats and tensor-core numerics for LLM inference (M2XFP, ASPLOS 2026; HPCA 2027).
Shanghai Jiao Tong University (SJTU)
Sep 2023 – Present
Undergraduate Student, ACM Honors Class
GPA: 3.90/4.3 (Rank: 11/27)
Academic Advisor: Prof. Yong Yu

Awards and Honors
  • 2025–2026, 2024–2025 & 2023–2024: Zhiyuan Honorary Scholarship

This homepage is adapted from a previous version of Xingyang Li's website, which itself was inspired by the designs of Haoran Geng and Jon Barron. And the background animation is from canvas-nest.js by hustcc
Last updated: Nov 3, 2026
© 2025 Zihan Zhang