Hi, I'm Du Jiaju.
I work on the layers between an agent and the GPU.
wuli666 on GitHub. I contribute to vLLM-Omni, own the worker refactor in vllm-rlt, build the desktop agent OpenDazi, and train GUI agents with RL.
Live trace
A simulated engine serving two GUI agents and background traffic, played at 1/20 speed.
- Throughput
- 0 tok/s
- Batch
- 0
- TTFT
- –
- TPOT
- –
- KV cache
- 0%
- MTP accept
- off
collectiveshared prefixpreemptionswitch flippedHover any bar for details.
About
I’m Du Jiaju, an engineer working on LLM inference. Most of my time goes into vLLM-Omni: I integrated Step-Audio2 there, and I fix bugs in duplex sessions, multi-stage pipelines and quantization.
I started out in computer vision. I trained detection and segmentation models for industrial inspection and traffic tracking, co-authored two papers on crop disease recognition, and deployed detection models to edge devices with TensorRT. From there I moved to LLM inference.
Right now I’m moving from integrating models to making them fast: generalizing eager MTP for TTS models, refactoring the worker and model runner of vllm-rlt, and learning to write kernels of my own. On the agent side, I build OpenDazi, a desktop AI agent for local files and meetings, and train Android GUI agents with reinforcement learning.
This site is my notebook. When I understand something well enough to explain it, it goes on the blog.
- Now
- LLM inference, GPU kernels, agents
How I got here
Computer science at Gansu Agricultural University.
First research: lightweight CNNs for crop disease recognition.
Computer vision engineer: inspection, tracking, TensorRT on edge devices.
langextract-vllm, a vLLM provider for LangExtract.
Step-Audio2, my first PR to vLLM-Omni.
OpenDazi, semantic-router working groups, and the vllm-rlt worker refactor.
Eager MTP for TTS models, and writing kernels.
Blog
Nothing published yet. Add a Markdown file to src/content/blog/ to write the first post.