I trained as a mathematician (MSc) and now build the infrastructure that large language models run on: training and inference frameworks, high-performance computing, and the CUDA kernels underneath. I like the problems where the profiler disagrees with the whiteboard — a kernel that should be faster, a pipeline that stalls, a bound that ought to be tight.
On the bench right now: LLM inference optimization, CUDA programming, algorithms. The through-line is depth in performance and systems work.
01 Yocto-Infra CUDA
Research notes, implementations and benchmarks on LLM inference optimization.
02 Yocto-AlgoGym C++
An algorithm training gym with its own test harness: implement, validate, optimize, repeat.
03 minimal-transformer Python
The smallest transformer I could still learn from.
04 UTGen C++
Unit-test generation tooling; the second iteration lives in UTGen-V2.
05 LCBench-Local Python
Run an LLM coding benchmark on your own machine.
Also on the shelf: benchmark. Upstreams I study: PaddleNLP · PaddleFormers · flash‑attention · opencompass · AIInfra
Happy to talk inference systems, CUDA, or hiking. yoctoinch@gmail.com · github.com/YoctoHan · Beijing
