Yizhuo Zheng (郑怡卓)
 |
Senior Student Majoring Computation Mathmatics |
Introduction to the basics
Now, I'm a senior student majoring Computation Mathmatics in Peking University. Computation Mathematics is the field of applied mathematics, which focuses on the study of how mathematical and logical problems can be effectively solved by computers.
I have paticipated in several scientific research including Software Development, Sketch, the Large language Models and so on.
Research
Research interests
Artificial Intelligence (Including: Machine Learning; Large Language Model)
Matrix Theory and Numerical Analysis
Scientific research experience
Interns Supervised by Professor Yang Tong, 02.2024-02.2025
Use Elastic Sketch to improve the classic KLL algorithm to achieve better quantile estimation in streams
Model distributed training traffic in the network, using the gaps to perform some active network measurements (ping) to avoid conflict with training traffic
Interns Supervised by Professor Jinbiao Wu, 11.2024-present
The Uzawa method is improved by using the augmented Lagrangian method, thereby accelerating the solution speed of large-scale sparse saddle point problems.
Model distributed training traffic in the network, using the gaps to perform some active network measurements (ping) to avoid conflict with training traffic
Recent publications
ongoing
Education
Main Courses: Numerical Algebra, Numerical Analysis, Optimization method, Programming of AI, Mathematical Introduction to Machine Learning, Parallel Computation
Internship Experience
Finite Element Industrial Software and Numerical Analysis Laboratory Project, 05.2023-05.2024
Researched on the recommendation system for users' preferences for selecting formulas.
Conducted the code testing on the library of numerical integration functions and corrected the errors of numerical and code errors.
Gained a grasp of Trilinos and reproduced the numerical methods used by Trilinos.
Large Language Model Inference Optimization Group in Baichuan AI, 06.2024-10.2024
Use Triton tools to develop large model reasoning, including: constructing corresponding block data for different hardware architectures to accelerate parallel optimization, and finally achieve a result that is 5% faster than the Cublas library.
Use C++ to call out the assembly source file compiled by the Triton tool to complete the code merge.
Explore W4A8's quantization tools, and combine W4A8+W8A8 to perform different quantization optimizations on matrices of different sizes.
Mechanical Transmission Simulation Laboratory, Chongqing Institute of Big Data, Peking University, 11.2024-05.2025
Computational performance tests of large-scale sparse matrix solving on Baltamatica and Matlab, including saddle point and non-saddle point problems.
Study the underlying principles of Mumps-5.6.2.2, write User’s Guidebook, mark and annotate function call relationships, try to reconstruct the multi-wavefront direct method based on MUMPS (complete Cholesky decomposition and reconstruction in the first phase, and use a third-party library for matrix reordering).