<![CDATA[元直的博客]]>https://blog.csdn.net/gsjthxyzh-cnhttps://blog.csdn.net/<![CDATA[通过 ACM 论文模板学习 LaTeX 语法：从零到顶刊投稿]]>https://blog.csdn.net/gsjthxy/article/details/165363961https://blog.csdn.net/gsjthxy/article/details/165363961gsjthxyMon, 14 Sep 2026 22:24:39 +0800<![CDATA[DFT 系列（03）：ATPG、故障模型与 Coverage——别被 99% 这个数字骗了]]>https://blog.csdn.net/gsjthxy/article/details/164326990https://blog.csdn.net/gsjthxy/article/details/164326990gsjthxyThu, 03 Sep 2026 10:29:36 +0800<![CDATA[DFT 系列（02）：Scan Cell、Scan Chain 与 Shift/Capture 协议]]>https://blog.csdn.net/gsjthxy/article/details/164326901https://blog.csdn.net/gsjthxy/article/details/164326901gsjthxyThu, 03 Sep 2026 10:27:53 +0800<![CDATA[DFT 系列（01）：先别急着插 Scan——DFT 的目标、架构与职责边界]]>https://blog.csdn.net/gsjthxy/article/details/164326767https://blog.csdn.net/gsjthxy/article/details/164326767gsjthxyThu, 03 Sep 2026 10:26:15 +0800<![CDATA[12 前置知识：PyTorch 速览（张量、自动微分与模块）]]>https://blog.csdn.net/gsjthxy/article/details/164301641https://blog.csdn.net/gsjthxy/article/details/164301641gsjthxyWed, 02 Sep 2026 17:23:53 +0800<![CDATA[11 训练进阶：学习率预热、余弦衰减与梯度裁剪]]>https://blog.csdn.net/gsjthxy/article/details/164301488https://blog.csdn.net/gsjthxy/article/details/164301488gsjthxyWed, 02 Sep 2026 17:19:26 +0800<![CDATA[10 高效微调：LoRA 实战（把参数量压到 1%）]]>https://blog.csdn.net/gsjthxy/article/details/164301369https://blog.csdn.net/gsjthxy/article/details/164301369gsjthxyWed, 02 Sep 2026 17:14:48 +0800<![CDATA[09 指令微调：训练一个能听懂指令的 AI 助手]]>https://blog.csdn.net/gsjthxy/article/details/164294060https://blog.csdn.net/gsjthxy/article/details/164294060gsjthxyWed, 02 Sep 2026 11:32:14 +0800<![CDATA[08 分类微调：把 GPT 变成垃圾邮件分类器]]>https://blog.csdn.net/gsjthxy/article/details/164289318https://blog.csdn.net/gsjthxy/article/details/164289318gsjthxyWed, 02 Sep 2026 10:41:22 +0800<![CDATA[07 预训练（下）：温度采样、Top-k 与加载官方 GPT-2 权重]]>https://blog.csdn.net/gsjthxy/article/details/164289220https://blog.csdn.net/gsjthxy/article/details/164289220gsjthxyWed, 02 Sep 2026 09:42:57 +0800<![CDATA[06 预训练（上）：损失、困惑度与训练循环]]>https://blog.csdn.net/gsjthxy/article/details/164264702https://blog.csdn.net/gsjthxy/article/details/164264702gsjthxyTue, 01 Sep 2026 17:45:38 +0800<![CDATA[05 从零搭建 GPT 模型：LayerNorm、GELU 与解码器堆叠]]>https://blog.csdn.net/gsjthxy/article/details/164260922https://blog.csdn.net/gsjthxy/article/details/164260922gsjthxyTue, 01 Sep 2026 15:13:56 +0800<![CDATA[04 注意力机制（下）：因果掩码与多头注意力]]>https://blog.csdn.net/gsjthxy/article/details/164260720https://blog.csdn.net/gsjthxy/article/details/164260720gsjthxyTue, 01 Sep 2026 15:06:15 +0800<![CDATA[03 注意力机制（上）：从 RNN 困境到自注意力与缩放点积注意力]]>https://blog.csdn.net/gsjthxy/article/details/164260575https://blog.csdn.net/gsjthxy/article/details/164260575gsjthxyTue, 01 Sep 2026 15:02:16 +0800<![CDATA[02 文本数据处理：分词、BPE 与词嵌入——给 LLM 准备“燃料”]]>https://blog.csdn.net/gsjthxy/article/details/164223288https://blog.csdn.net/gsjthxy/article/details/164223288gsjthxyMon, 31 Aug 2026 16:45:00 +0800和<|endoftext|>等特殊token处理未知词与文档边界。最后强调嵌入向量将语义相近的词映射到相近位置，是模型理解语言的基础。]]><![CDATA[01 理解大模型：为什么从零构建？——LLM 全景与学习路线图]]>https://blog.csdn.net/gsjthxy/article/details/164216890https://blog.csdn.net/gsjthxy/article/details/164216890gsjthxyMon, 31 Aug 2026 15:10:02 +0800<![CDATA[00 工程篇：环境搭建与项目结构（从零开发 GPT 的开工准备）]]>https://blog.csdn.net/gsjthxy/article/details/164216460https://blog.csdn.net/gsjthxy/article/details/164216460gsjthxyMon, 31 Aug 2026 15:01:20 +0800<![CDATA[DFT 设计与验证系列（专业版）：从 Scan 到 SoC/Chiplet Sign-off]]>https://blog.csdn.net/gsjthxy/article/details/164117700https://blog.csdn.net/gsjthxy/article/details/164117700gsjthxyThu, 27 Aug 2026 14:38:20 +0800<![CDATA[DFT 系列（14）：DE/DV 项目落地——Milestone、Testplan、Case 与 Sign-off 模板]]>https://blog.csdn.net/gsjthxy/article/details/164109710https://blog.csdn.net/gsjthxy/article/details/164109710gsjthxyThu, 27 Aug 2026 09:42:50 +0800<![CDATA[DFT 系列（13）：Coverage、Sign-off 与 Debug——先分类，再看波形]]>https://blog.csdn.net/gsjthxy/article/details/164109661https://blog.csdn.net/gsjthxy/article/details/164109661gsjthxyThu, 27 Aug 2026 09:41:17 +0800