Dayou Du

Phd of Computer Science

University of Edinburgh

Biography

My name is Dayou Du (杜大猷). I am currently a first-year PhD student in Computer Science at the University of Edinburgh, advised by Dr. Jianyi Cheng and Dr. Luo Mai. Previously, I earned my MPhil degree from HKUST(GZ), where I was advised by Prof. Xiaowen Chu, and my B.Eng. from South China University of Technology. During the summers of 2023 and 2024, I interned at Microsoft Research Asia, working with Shijie Cao.

Interests

Efficient Deep Learning Systems
Hardware Accelerators

Education

Phd in Computing Systems Architecture, 2025

The University of Edinburgh
MPhil in Data Science and Analytics, 2022

The Hong Kong University of Science and Technology
BE in Intelligent Science and Technology, 2018

South China University of Technology

Experience

Research Intern

Microsoft Research Asia

July 2024 – March 2025 Beijing

Research on long-context LLMs inference.

Research Intern

Microsoft Research Asia

July 2023 – December 2023 Beijing

Research on model quantization of LLMs.

Engineering Intern

June 2021 – October 2021 GuangZhou

Designed a framework for Nature Language Understanding tasks.

Publications

BitDecoding: Unlocking Tensor Cores for Long-Context LLMs Decoding with Low-Bit KV Cache

BitDecoding is a GPU-optimized framework that enhances decoding efficiency for long-context LLMs by unlocking Tensor Cores with low-bit KV cache, achieving significant speedups and reducing latency in long-context generation scenarios.

Dayou Du, Shijie Cao, Jianyi Cheng, Ting Cao, Mao Yang

BitDistiller: Unleashing the Potential of Sub-4-Bit LLMs via Self-Distillation

[ACL 2024] BitDistiller synergizes Quantization-Aware Training and Knowledge Distillation, significantly enhancing sub-4-bit LLMs’ performance on complex tasks with reduced data and training resource demands.

Dayou Du, Yijia Zhang, Shijie Cao, Jiaqi Guo, Ting Cao, Xiaowen Chu, Ningyi Xu

SeerAttention: Learning Intrinsic Sparse Attention in Your LLMs

This study introduces SeerAttention, an attention mechanism that learns block-level sparsity directly from LLMs, enhancing efficiency and scalability in long-context processing without relying on predefined patterns.

Yizhao Gao, Zhichen Zeng, Dayou Du, Shijie Cao, Peiyuan Zhou, Jiaxing Qi, Junjie Lai, Hayden Kwok-Hay So, Ting Cao, Fan Yang, Mao Yang

Benchmarking and Dissecting the Nvidia Hopper GPU Architecture

[IPDPS 2024] This research benchmarks Hopper GPUs, analyzing new features like FP8 tensor cores and DPX, to aid in software optimization and understanding of their microarchitectural and programming nuances.

Weile Luo, Ruibo Fan, Zeyu Li, Dayou Du, Qiang Wang, Xiaowen Chu