Resume

No. 100, Fuxing Rd., High-Tech District, Hefei, Anhui Province, China 230031
 sh.fu@outlook.com ⬦  Fr4nk1inCs

You can get a PDF version here.

Research Interests

LLM inference optimization, System for MoE.

DeaMoE: Efficient MoE Architecture for Fast Small Batch DecodingADSL, USTC
Core MemberDec 2025—Jan 2026

A novel MoE architecture using expert grouping and parameter sharing to significantly reduce expert weight loading per decoding step for small-batch workloads while preserving accuracy.
Responsible for the adaptation of DeaMoE architecture in the training framework Megatron-LM and the inference framework vLLM.
Developed custom inference operators for two-stage Routing strategy, achieving decoding speedup proportional to weight loading reduction on A40 and H100 GPUs.

Parallelism Planning for MoE Inference with Dynamic Top-K RoutingADSL, USTC
Core MemberMar 2025—Aug 2025

An inference framework for dymamic top-k routing MoE models, which automatically plans parallelism strategies to maximize throughput on prefill-dominated workloads.
Paricipated in the implementation of the model profiler, adoption of dynamic top-k routing, pipeline parallelism enhancements, and the design of the parallelism planner.

[1] Zewen Jin, Shen Fu, Chengjie Tang, Youhui Bai, Shengnan Wang, Jiaan Zhu, Chizheng Fang, Ping Gong, and Cheng Li. 2026. SMIDT: High-Performance Inference Framework for MoE Models with Dynamic Top-K Routing. Proceedings of the AAAI Conference on Artificial Intelligence 40, 27 (March 2026), 22444–22453. https://doi.org/10.1609/aaai.v40i27.39403

University of Science and Technology of ChinaHefei, Anhui
M.E. in Computer Science and TechnologySep 2024—Present

University of Science and Technology of ChinaHefei, Anhui
B.E. in Computer Science and TechnologySep 2020—Jun 2024

Services

Teaching

T.A. for Compiler Principles and Techniques (Instructor: Prof. Cheng Li)2023 Autumn, USTC

Open Source Contributions

Skills