张逸骅的博客 | Yihua's Blog

From GRPO to DAPO and GSPO: What, Why and How [En/中]

GRPO 的进化之路：从 GRPO 走向 DAPO 和 GSPO

Posted by Yihua Zhang on August 8, 2025

Re-understanding KL Approximation from an RL-for-LLM Lens: Notes on “Approximating KL Divergence [En/中]

从 RL for LLM 视角重新理解 KL 估计：读《Approximating KL Divergence》笔记

Posted by Yihua Zhang on July 2, 2025

Decorators in Machine Learning Projects [En/中]

机器学习中的装饰器

Posted by Yihua Zhang on April 10, 2025

《Bauklötze》音乐解构

积木崩塌时的命运回响，泽野弘之用音符砌筑的巨人悲歌

《Bauklötze》音乐解构 - 积木崩塌时的命运回响，泽野弘之用音符砌筑的巨人悲歌想为《进击的巨人》的 OST 写音乐鉴赏，已经是我心里惦记很久的事了。一方面，泽野弘之的配乐确实令人动容；另一方面，《进击的巨人》这部作品与泽野的音乐在情感与氛围的结合上堪称 1+1 > 2，每当音符响起，观众的心弦也随之颤动。于是，我在这里开个新坑，把自己最喜欢的一系列配乐做成音乐鉴赏，算是我读...

Posted by Yihua Zhang on March 8, 2025

DualPipe Explained: A Comprehensive Guide to DualPipe That Anyone Can Understand—Even Without a Distributed Background [En/中]

DualPipe 深入浅出：没有分布式训练基础也能看懂的 DualPipe 全方位讲解

Posted by Yihua Zhang on February 27, 2025

Navigating the RLHF Landscape: From Policy Gradients to PPO, GAE, and DPO for LLM Alignment [En/中]

大语言模型 RLHF 全链路揭秘：从策略梯度、PPO、GAE 到 DPO 的实战指南

Posted by Yihua Zhang on February 11, 2025

DeepSeek-R1 Dissection: Understanding PPO & GRPO Without Any Prior Reinforcement Learning Knowledge [En/中]

DeepSeek-R1 技术剖析：没有强化学习基础也能看懂的 PPO & GRPO

Posted by Yihua Zhang on February 7, 2025

Why Cache 32 Heads When One Latent Variable Suffices? A Theory-to-Code Guide to DeepSeek’s MLA for KV-Cache [En/中]

从多头共享到潜变量：DeepSeek的MLA在低秩投影与按需解压中重新定义 KV-Cache

Posted by Yihua Zhang on February 2, 2025

From Zero to Reasoning Hero: How DeepSeek-R1 Leverages Reinforcement Learning to Master Complex Reasoning [En/中]

千呼万唤始出来：DeepSeek-R1 如何通过强化学习实现复杂推理

Posted by Yihua Zhang on January 20, 2025

A Review on the Evolvement of Load Balancing Strategy in MoE LLMs: Pitfalls and Lessons [En/中]

关于 MoE 大模型负载均衡策略演进的回顾：坑点与经验教训

Posted by Yihua Zhang on January 15, 2025

ABOUT ME

Yihua's Blog