llm-lite 는 저사양 로컬 환경에서 Gemma 3N E4B 를 클라우드 없이 돌리는 걸 목표로 만든 멀티 백엔드 추론 엔진이다. 모델 구조는 그대로 두되 공격적인 양자화(INT4 weights + MMAP)와 저수준 하드웨어 가속으로 성능을 끌어내는 방향을 택했다.

This text contains the core concepts and mathematical principles of the Transformer model architecture.

This document is a note organizing the architecture and training process of the GPT-1 paper by combining mathematical definitions with intuitive interpretations.

Cherry blossoms are starting to bloom — and so is this new site.

I finally set up a proper personal homepage. What used to be just a GitHub Profile README is now a full static site powered by Docusaurus.

A quick summary of notable updates in deep learning inference, GPU architecture, and HPC this week.

The book that helped me most when I first started learning CUDA programming.

While studying deep learning inference optimization, I explored how memory access patterns in CUDA kernels dramatically affect performance.