CV - Lihui Gu(顾立辉)/Academic Pages

Education

M.S. in Computer Science, Zhejiang University, 2023 - 2026(Expected)

B.S. in Computer Science and Technology, Zhejiang University of Technology, 2019 - 2023

Computer Aided Design and Graphics System National Key Laboratory, Zhejiang University
School of Computer Science and Technology, Zhejiang University of Technology
Double Major: Computer Science and Technology; Minor: Intelligent Science and Technology
GPA: 4.23/5, Major Ranking: 1/23

Work experience

Large Model Inference Optimization Intern, Kuaishou - Keling Technology Department, Beijing 2025.06 - 2025.09:

Optimized feature cache reuse strategy for DiT models by leveraging the similarity of adjacent denoising time-step features. Achieved a 2.5x end-to-end acceleration on the Wan2.1/Hunyuan video generation model and 3.0x on the Flux image generation model, surpassing SOTA solutions like Taylorseer/Easycache under the same acceleration ratio. Nearly lossless 1.4x end-to-end acceleration verified on the un-distilled Keling i2v model.

Navigation Group Algorithm Intern, Hangzhou Feibu Technology, Zhejiang, 2024.09-2025.04

Predictive-Interactive Integrated Modeling: Designed an end-to-end data-driven model to replace rule-based strategies, enhancing decision-making in complex scenarios.
Reinforcement Learning Optimization: Implemented model iterations based on the PPO algorithm, continuously improving port intersection traffic efficiency and reducing collision risks.
Engineering Implementation: Reduced average intersection passage time by 0.5s across 1000+ simulation cases, completed real vehicle deployment of the model, and received an S-level performance evaluation during the internship.

Skills

Deep Learning & Inference Optimization:

Familiar with DIT/LLM inference optimization techniques, core functionalities of mainstream inference frameworks (vLLM, xDiT), and multi-card distributed inference technology.
Skilled in the underlying implementation of speculative decoding, combining FP8 quantization, communication optimization, tensor parallelism, and CUDA Graph techniques.
Knowledgeable in the implementation details of high-performance fusion operators (Megakernel), achieving extreme acceleration by fusing the entire LLM decoding process into one kernel.

Industrial Project Experience:

Led deep learning/reinforcement learning projects in autonomous driving scenarios (prediction-interaction tasks), with comprehensive experience from algorithm to engineering.

Publications