ZEYU WANG
341-***-****· ad4cij@r.postjobfree.com *** E Duane Ave, Sunnyvale, CA 94085
EDUCATION
UNIVERSITY OF SOUTHERN CALIFORNIA Sep 2022- Jun 2024 MS in Computer Engineering
XIAMEN UNIVERSITY Sep 2018- Jun 2022
BS in Computer Science and Technology
EXPERIENCE
TENCENT TECHNOLOGY(BEIJING) COMPANY Jan 2022- May 2022 Intern, Hotspot Application Group, News algorithm research and development department, Tencent news
Built a ETL pipeline for hotspot recommendation and recall channel of hot events.
Used RPC to request remote services in a distributed system, to integrate the content of an event for recommendation algorithm team.
Designed and implemented core sorting logic in both C++ and Python, including intervention rules and cosine angle similarity calculation.
Applied optimized algorithm to reduce the total running time from 15ms to 3ms (C++), integrated with the middle platform of Tencent News.
Implemented real-time news browsing data stream analysis using Spark streaming to help locating potential hot pot.
Using Hadoop HDFS to store news browsing data, reduced batch processing time for MapReduce by 18%.
Built news browsing data integrity verification function using Spring Boot and Oracle Database.
Developed real-time news browsing data visualization using React and E-Charts. PROJECT
Efficient Fine-tuning of LLM on a Single GPU
Implement the inference process for the LLaMA2 7B model. Implement LoRA Linear Module and convert the model into PEFT (Parameter-efficient Fine-tuning) model by replacing Q, V projection layers into LoRA linear modules, reducing 99% memory usage. Used Pytorch’s AMP API to modify the training loop to enable FP16 mixed precision training. Trained the LLaMA2 7B model on the Alpaca Dataset. Reduced 99.88% training parameters, from 6.74b to 8.39m. Only 35821MiB memory was used after using all the techs into training. Efficient Distributed Training
Applied single shot network and early bird tickets pruning methods to the FedAvg framework on the server and client side, reduced the parameters to 50% of the original. Quantized weights and activations of a CNN network and reduced 40% memory usage with same accuracy. Used MPI to implement various types of distributed training methods including parameter-sever and federated learning on non-IID and IID distribution datasets. Analyzed FedAudio and PipeTransformer framework and analyzed the parameters and time complexity on tree-based and ring-based reduction. Health Status Management System
Developed backend REST API using SpringBoot to manage user personal information and health status.
Implemented MySQL database to store user data, and added MyBatis persistence layer to simplify database access.
Designed front-end management interfaces using React and visualized status summary using Google Charts. Smart Farm Monitoring System
Built a Self-Sufficient IoT device using MultiTech xDot, forwarded sensor data to AWS IoT Core.
Using Node-RED and visualized sensor data using Things Board dashboard.
Deployed Flask backend server on AWS EC2, implemented CRUD function using boto3 with DynamoDB.
Implemented smart suggestions based on weather forecast using machine learning model and OpenWeather API.
Developed iOS application using SwiftUI to display real-time environment condition and smart suggestions. TECHNICAL SKILLS
Languages and Frameworks: Java, Python, C/C++, JavaScript, SQL, HTML/CSS, Swift, Spring Boot, Maven, Flask, Spark, Hadoop, React, MyBatis
Databases and Tools: MySQL, Oracle Database, Redis, DynamoDB, Linux, Git, Docker, AWS, Jira