Hey, I'm Robin 👋
RSS FeedExploring how AI models learn, scale, and serve.
I'm an engineer based in Shenzhen, working on LLM systems and reinforcement learning frameworks. My interests span post-training, inference optimization, and distributed training.
This is my personal corner of the web: technical deep dives, open-source projects, and reflections on work and life. For structured tutorials and research notes, visit Robin's AI Lab.
What I explore
Connecting algorithmic ideas with practical systems. Explore these topics in Robin's AI Lab.
-
Post-training & reinforcement learning
From RL fundamentals to RLHF and GRPO: understanding how learning objectives, data, and training systems fit together.
-
Efficient LLM inference
Attention algorithms, scheduling, and serving systems, with a focus on the ideas behind vLLM and SGLang.
-
Distributed training & AI hardware
Parallel training, MoE communication, and the trade-offs of bringing workloads to the Ascend ecosystem.
Selected open source
Implementations for exploring the ideas behind AI systems.
-
mini-vLLM
Explore the building blocks of modern LLM serving through a minimal implementation.
-
ScaleTorch
Study tensor, pipeline, sequence, and data parallelism in PyTorch.
-
Deep-RL-Toolkit
Explore single-agent reinforcement learning algorithms, including DQN, PPO, and SAC.
Featured
-
从 Online Softmax 到 FlashAttention
从数值稳定的 Safe Softmax 出发，推导 Online Softmax 的递推公式，最终理解 FlashAttention 如何将注意力计算融合为单轮遍历的 IO 感知算法。
-
KL散度的近似计算方法
探讨使用蒙特卡洛方法对KL散度进行近似估计的技术，分析三种估计量的偏差与方差特性，并推广至任意f-散度。
-
Trackio：Hugging Face 开源的免费实验追踪库，wandb 的即插即用替代方案
Trackio 是 Hugging Face 推出的开源免费 Python 实验追踪库，API 与 wandb 完全兼容，支持本地可视化仪表板和 Spaces 云端共享。
-
拒绝 GRPO 焦虑！离线强化学习也能造就数学最强基座？PCL-Reasoner-V1.5 技术深度解析
PCL-Reasoner-V1.5 基于离线强化学习在 Qwen2.5-32B 上实现 AIME 2024/2025 SOTA 成绩，探讨 Offline RL 作为 GRPO 替代方案的技术路径。