LLM Inference
From attention and quantization to CUDA kernels, I study and optimize the full LLM inference path.
AI SYSTEMS · HIGH-PERFORMANCE COMPUTING
Hi, I’m Zhao Dongyu.
I’m a programmer and engineer focused on building faster, smaller, and more reliable AI systems. My current interests include large language model inference, high-performance computing, edge AI, and embodied intelligence.
ABOUT
Engineering between algorithms and hardware.
I graduated from Peking University and have spent years studying and building AI systems, inference frameworks, and high-performance computing software. I enjoy connecting papers, source code, and hardware behavior—and using reproducible experiments to find where systems actually hit their limits.
Learn more about meFOCUS
Problems I care about.
High-Performance Computing
Mapping algorithms to GPU and RISC-V hardware through GEMM, memory hierarchies, and performance analysis.
Edge AI
Exploring TensorFlow Lite/Micro, on-device deployment, and model efficiency under tight resource constraints.
Embodied AI
Working with VLA models, robot simulation, and real-world deployment across the perception-to-action pipeline.
SELECTED WORK
Selected projects and technical writing.
Practicing and Understanding the π Model
A structured look at Physical Intelligence’s π models, from architecture to simulation and real-world deployment.
FlashMLA in Practice
An experimental analysis of MLA kernel source code, profiling data, and performance behavior.
Optimizing SGEMM on RISC-V
Improving matrix multiplication through cache locality, blocking, and RVV assembly.
Accelerating Workloads on an RTX 4090
Performance exploration across mixed precision, Tensor Cores, and custom operator implementations.
CONTACT
Let’s talk systems, projects, and collaboration.
If you’re working on AI systems, high-performance computing, or embodied intelligence, I’d be happy to connect.