<![CDATA[WitsMakeMen的专栏]]>https://blog.csdn.net/WitsMakeMenzh-cnhttps://blog.csdn.net/<![CDATA[RankMixer论文理解]]>https://blog.csdn.net/WitsMakeMen/article/details/159655264https://blog.csdn.net/WitsMakeMen/article/details/159655264WitsMakeMenMon, 30 Mar 2026 21:12:30 +0800<![CDATA[relu激活函数为什么有效]]>https://blog.csdn.net/WitsMakeMen/article/details/159654888https://blog.csdn.net/WitsMakeMen/article/details/159654888WitsMakeMenMon, 30 Mar 2026 20:24:32 +0800<![CDATA[rq-vae训练过程是什么样的]]>https://blog.csdn.net/WitsMakeMen/article/details/159395768https://blog.csdn.net/WitsMakeMen/article/details/159395768WitsMakeMenMon, 23 Mar 2026 21:50:40 +0800<![CDATA[PPO算法原理]]>https://blog.csdn.net/WitsMakeMen/article/details/159214224https://blog.csdn.net/WitsMakeMen/article/details/159214224WitsMakeMenWed, 18 Mar 2026 21:42:44 +0800<![CDATA[RoPE 算法原理？算法为什么只和相对位置有关]]>https://blog.csdn.net/WitsMakeMen/article/details/159009962https://blog.csdn.net/WitsMakeMen/article/details/159009962WitsMakeMenFri, 13 Mar 2026 15:06:34 +0800<![CDATA[slice作用]]>https://blog.csdn.net/WitsMakeMen/article/details/158976812https://blog.csdn.net/WitsMakeMen/article/details/158976812WitsMakeMenThu, 12 Mar 2026 20:20:48 +0800<![CDATA[rankerMixer为什么能提升系统的MFU]]>https://blog.csdn.net/WitsMakeMen/article/details/158934652https://blog.csdn.net/WitsMakeMen/article/details/158934652WitsMakeMenWed, 11 Mar 2026 21:41:19 +0800<![CDATA[优化器状态也放 CPU（CPUOffload）具体做法]]>https://blog.csdn.net/WitsMakeMen/article/details/158422124https://blog.csdn.net/WitsMakeMen/article/details/158422124WitsMakeMenThu, 26 Feb 2026 11:58:12 +0800<![CDATA[LLM基座发展动向（2026最新·详细版）]]>https://blog.csdn.net/WitsMakeMen/article/details/158039340https://blog.csdn.net/WitsMakeMen/article/details/158039340WitsMakeMenFri, 13 Feb 2026 14:07:55 +0800<![CDATA[10 年后 AGI 能否出现？将如何影响人类生活？]]>https://blog.csdn.net/WitsMakeMen/article/details/157221561https://blog.csdn.net/WitsMakeMen/article/details/157221561WitsMakeMenWed, 21 Jan 2026 19:49:24 +0800<![CDATA[Qwen 3 技术解析：通过强化学习和思维模式融合提升模型对齐]]>https://blog.csdn.net/WitsMakeMen/article/details/157215335https://blog.csdn.net/WitsMakeMen/article/details/157215335WitsMakeMenWed, 21 Jan 2026 15:11:25 +0800<![CDATA[用矩阵实例具象化 RankMixer 核心机制]]>https://blog.csdn.net/WitsMakeMen/article/details/156954388https://blog.csdn.net/WitsMakeMen/article/details/156954388WitsMakeMenWed, 14 Jan 2026 21:58:44 +0800<![CDATA[BLEU计算原理]]>https://blog.csdn.net/WitsMakeMen/article/details/156913779https://blog.csdn.net/WitsMakeMen/article/details/156913779WitsMakeMenTue, 13 Jan 2026 21:48:56 +0800<![CDATA[理解ppl指标]]>https://blog.csdn.net/WitsMakeMen/article/details/156913318https://blog.csdn.net/WitsMakeMen/article/details/156913318WitsMakeMenTue, 13 Jan 2026 21:20:42 +0800<![CDATA[SLMRec 论文中解决Dnorm​损失多层收敛不稳定的「核心 3 个技术方案」（论文原文答案，全部是技术亮点）]]>https://blog.csdn.net/WitsMakeMen/article/details/156690440https://blog.csdn.net/WitsMakeMen/article/details/156690440WitsMakeMenWed, 07 Jan 2026 16:02:22 +0800<![CDATA[SLMRec 详细介绍（重点：损失函数）]]>https://blog.csdn.net/WitsMakeMen/article/details/156570843https://blog.csdn.net/WitsMakeMen/article/details/156570843WitsMakeMenSun, 04 Jan 2026 15:56:10 +0800<![CDATA[qwen3 训练loss 出现nan]]>https://blog.csdn.net/WitsMakeMen/article/details/156397747https://blog.csdn.net/WitsMakeMen/article/details/156397747WitsMakeMenMon, 29 Dec 2025 21:06:02 +0800<![CDATA[训练时开启 KV 缓存会和is_causal=False 冲突]]>https://blog.csdn.net/WitsMakeMen/article/details/156393118https://blog.csdn.net/WitsMakeMen/article/details/156393118WitsMakeMenMon, 29 Dec 2025 17:09:39 +0800<![CDATA[大语言Qwen3模型QA训练数据如何生成]]>https://blog.csdn.net/WitsMakeMen/article/details/156386678https://blog.csdn.net/WitsMakeMen/article/details/156386678WitsMakeMenMon, 29 Dec 2025 12:07:09 +0800<![CDATA[scaled_dot_product_attention实现]]>https://blog.csdn.net/WitsMakeMen/article/details/156309492https://blog.csdn.net/WitsMakeMen/article/details/156309492WitsMakeMenFri, 26 Dec 2025 16:43:03 +0800