Systems and LLM Inference Engineering
I'm Jinyang Su — LLM inference engineer, formerly database storage engineer. Here I write about inference engines, CUDA kernels, RDMA, and storage internals — mostly notes to my future self. More about me.
Posts
- DeepSeek V4.1 推理笔记：KV Cache、Engram 与 DSpark
- Understanding Deep Learning Requires Rethinking Generalization
- 从 cubin 到 kern：引擎对模型的感知面可以压缩到什么程度
- When LLMs Flip Coins, Is the Probability 50%?
- pegainfer (6): MoE Expert Parallelism From Scratch
- pegainfer (7): MoE Expert Parallelism From Scratch (Part 2)
- cuTile Intro
- pegainfer (5): One Size Can't Fit All
- How I Vibe Code
- Pegaflow (1): RDMA From Zero to Data Transfer
- pegainfer (4): From Pre-allocation to Graph Replay
- pegainfer (3): From Launch Overhead to CUDA Graph (Part 1)
- pegainfer (2): Adding a Sampler to the Inference Engine
- pegainfer: A Native Rust Inference Engine from Scratch