🎵About Me
I am a third-year Ph.D. student in Computer Science and Engineering at the University of California San Diego, working in the STABLE lab with Prof. Jishen Zhao. My research spans machine learning systems, high-performance computing, and computer architecture, with a focus on efficient LLM inference.
I develop algorithms, GPU kernels, and runtimes for sparse attention, mixture-of-experts prefill, and heterogeneous acceleration. My recent work includes collaborations with Qualcomm AI Research and TPU optimization at Google in Summer 2026. Previously, I received my B.S. from the ACM Class at Zhiyuan College, Shanghai Jiao Tong University.
- Machine Learning Systems
- High-Performance Computing
- Computer Architecture
Ph.D. in Computer Science and Engineering
University of California San Diego, CSE
B.S. in Computer Science and Technology
Shanghai Jiao Tong University, ACM Class, Zhiyuan College
I work on making LLM inference faster and more resource-efficient through algorithm and system co-design.
- Sparse attention on mobile SoCs: TileWeave combines NPU/GPU execution with a coordinated llama.cpp runtime, achieving 1.35–3.23× faster full-model prefill than XAttention on Snapdragon phones.
- Mixture-of-experts inference: RankLane uses load-ranked expert capacities and locality-preserving aggregation to accelerate Qwen3 MoE prefill by 1.27–4.18× versus a dense baseline on Cloud AI 100 Ultra.
- TPU optimization: At Google in Summer 2026, I benchmarked 23 TPU7x optimization tasks and developed hierarchical agent-driven optimization, achieving 2.86× geometric mean speedup on unseen shapes and 1.62× end-to-end serving throughput.
- Long-context LLMs: Double-P uses hierarchical top-p sparse attention, adaptive token budgets, and GPU kernels to reduce selection overhead while preserving attention mass.
Please reach out to collaborate 😃
TileWeave: Heterogeneous Tiling for Sparse Prefill on Mobile SoCs
Wentao Ni, Hong Cai, Mingu Lee, Chihao Yu, Zhongming Yu, Hejia Zhang, Kangqi Zhang, Fatih Porikli, Zhijian Liu, Jishen Zhao. Submitted to MLSys 2027.
RankLane: Efficient MoE Prefill on Accelerators with Ranked Expert Capacities
Wentao Ni, Chihao Yu, Hong Cai, Mingu Lee, Hejia Zhang, Fatih Porikli, Jishen Zhao. Submitted to MLSys 2027.
TIERBENCH and TIERFLOW: Benchmarking and Stabilizing Multi-Tier Optimization on TPU7x
Submitted to MLSys 2027. Work completed at Google, Summer 2026.
Double-P: Hierarchical Top-P Sparse Attention for Long-Context LLMs
Wentao Ni, Kangqi Zhang, Zhongming Yu, Oren Nelson, Mingu Lee, Hong Cai, Fatih Porikli, Jongryool Kim, Zhijian Liu, Jishen Zhao. Submitted to NeurIPS 2026 (under review).