https://auto-driving-blog.pages.dev/tags/%E6%A2%AF%E5%BA%A6%E6%B5%81/ Recent content in 梯度流 on Elon's AD Insight https://auto-driving-blog.pages.dev/images/share.png https://auto-driving-blog.pages.dev/images/share.png Hugo zh-cn Thu, 30 Jul 2026 00:00:00 +0000 https://auto-driving-blog.pages.dev/posts/thoughts/flow-grpo-complete-guide/ Thu, 30 Jul 2026 00:00:00 +0000 https://auto-driving-blog.pages.dev/posts/thoughts/flow-grpo-complete-guide/ 从 train_flux_fast.py 第 1 行开始，逐层追踪 Flow-GRPO 的完整逻辑链：采样阶段做了什么？reward 怎么变成 advantage？训练阶段的计算图是怎么构造的？loss 为什么那样设计？梯度如何从最后一个 log_prob 传到 LoRA 参数？每段代码都标注了源文件行号。