ShadmanRohan

🎓 CS grad | 🧠 Computational Linguistics & Vision
he/him
Linkedin:

[Jason Zhu @GoSailGlobal]

Stanford CS336 上，Tatsu 讲了一节 LLM 架构课，把过去 3 年所有主流 LLM 拆开，看它们的共通模板

结论挺爆：90% 的架构选择已经收敛，你随便挑一个开源大模型，它跟其他模型在这些维度上几乎一模一样

讲师的原话
- 2024 年大家都在 cosplay Llama2
- 2025 年的主题是「怎么训得不崩」
- 2026 年的主题是「怎么扛住长上下文」

下面是 2026 年开源 LLM 的标准模板 你训自己的模型可以直接抄

【架构层 已经收敛的 7 件事】
1）Layer Norm 挪出残差流（pre-norm）
原版 Transformer 把 LN 放在残差里 几乎所有现代模型都挪到外面
原因：keep your residual stream clean 梯度反传更稳

2）RMS Norm 替代 LayerNorm
LayerNorm 的减均值 + 加 bias 那部分实际没怎么帮上忙
丢掉之后 flops 只省 0.17% 但运行时省到 25%
（瓶颈在数据搬运 计算反而次要）

3）所有 bias 项全删
跟 RMS Norm 一个道理 系统层省内存搬运

4）激活函数用 SwiGLU 或 GeGLU
gated linear unit 几乎所有现代模型都用
Llama 系 / Qwen / Mistral 用 SwiGLU
Google 系（Gemma / T5）用 GeGLU
区别极小 选哪个都行

5）位置编码用 RoPE
2024 年之后基本统一了
原理：把每对维度按位置旋转一个角度 让 inner product 只依赖相对位置

6）Transformer block 串联（不是并联）
GPT-J / Palm 试过并联 现在基本被放弃
串联的实现优化得太好了 并联省的那点系统开销不值得损失表达力

7）Layer norm 可以「撒」
哪儿不稳就在哪儿加 LN
attention 之前能加 之后能加 两边都加（double norm）也可以
现代模型很多这样做

【超参数 已经收敛的 5 个数】
1）feedforward 维度 / hidden 维度
- 非 GLU 模型：4 倍
- GLU 模型：8/3 ≈ 2.67 倍（因为 GLU 多一组矩阵 要保持总参数量）
- Llama 系：3.5 倍
- T5 1.0 试过 64 倍 后来 T5 1.1 改回标准 别学

2）head 数 × head 维度 ≈ hidden 维度
几乎所有模型都遵守 T5 是为数不多的例外

3）模型纵横比（hidden / 层数）≈ 100
太深 pipeline parallel 难做
太宽 表达力受限
100 这个数字是系统约束 + 表达力的平衡点

4）vocab size
单语模型：30K 左右（早期 GPT-2 那种）
多语 / 通用模型：100K-200K（GPT-4 / Llama 3 / Gemma 都在这个范围）
现代基本都是后者

5）weight decay
仍然普遍使用
但研究发现它在 LLM 里干的事其实是优化器干预 让你最终能收敛到更深的最优点
跟你想的「防过拟合」没什么关系
所以别因为「单 epoch 不会过拟合」就把它关掉

【稳定性 三个救命 trick】
训练大模型最怕中途 loss 突然飙升 然后 NaN 全军覆没
现代模型用三个 trick 防这件事

1）Z-loss
output softmax 的 normalizer 容易爆
加一个 (log Z)² 的正则项 让 Z 始终接近 1
DCLM / Olmo 都用

2）QK norm
attention 的 Q 和 K 在矩阵乘之前各加一个 LN
让 softmax 的输入永远是单位尺度
multimodal 圈先用起来 现在所有大模型都加

3）Logit soft cap（仅 Google 系）
attention logit 用 tanh 硬封顶
Gemma 2/3/4 都在用 但会损失一点点性能 慎用

【Attention 两个新趋势】
1）GQA（Grouped Query Attention）几乎统一
原版 multi-head 推理时 KV cache 会让算术强度崩到 1/h
GQA 共享 K 和 V 但保留多个 Q
表达力几乎不损失 推理成本砍掉 80%
现在所有要做生产部署的大模型 没有不用 GQA 的

2）局部 + 全局 attention 交替
处理长上下文的新方式
Cohere Command A 起头 现在 Llama 4 / Gemma 4 / Olmo 3 全在用
比如每 4 层有 1 层 full attention 其他 3 层是 sliding window 只看附近的 token
比纯 SSM 更稳 比纯 full attention 便宜得多
（Qwen 3.5 做了变体 把 sliding window 那 3 层换成 SSM）

收尾一句
如果你正在训自己的 LLM，上面这一套就是 2026 年的「默认配置」 不需要重新发明，直接抄

如果你只是想看懂 GitHub 上那些 modeling_xxx.py
这一份足够你不再被术语吓住

[Roan @RohOnChain]

This 2 hour Stanford lecture shows exactly how Stanford trains it's engineers to build AI systems. It's more practical than every Claude tutorial & prompting threads you've seen.

Bookmark & give it 2 hours, no matter what. It'll be the most productive thing you do this weekend.

[François Chollet @fchollet]

AI agents will soon graduate to fully-fledged economic actors that buy services, compute, and even data in the course of accomplishing high-level goals. 1-2 years before we start seeing this at scale.

[Boris Cherny @bcherny]

I'm Boris and I created Claude Code. Lots of people have asked how I use Claude Code, so I wanted to show off my setup a bit.

My setup might be surprisingly vanilla! Claude Code works great out of the box, so I personally don't customize it much. There is no one correct way to use Claude Code: we intentionally build it in a way that you can use it, customize it, and hack it however you like. Each person on the Claude Code team uses it very differently.

So, here goes.

[ShadmanRohan @RohanShadman]

2019 vs. Today. We’ve come a long way.

Back then, the “gotcha” was: ask it a simple arithmetic word problem and it collapses. Today, Fields Medalists are using these models to turn research math into machine-checkable proofs.

[Akshay 🚀 @akshay_pachaar]

When outputs are verifiable, labels become optional.

Maths, code, and logic can be automatically checked and validated.

Let's use this fact to build a reasoning model without manual labelling.

We'll use:

- @UnslothAI for parameter-efficient finetuning.
- @HuggingFace TRL to apply GRPO.

Let's go! 🚀

[Paata Ivanisvili @PI010101]

GPT-5 Pro found a counterexample to the NICD-with-erasures majority optimality (Simons list, p.25).
https://t.co/T3m9MYgqe0

At p=0.4, n=5, f(x) = sign(x_1-3x_2+x_3-x_4+3x_5) gives E|f(x)|=0.43024 vs best majority 0.42904.

[ShadmanRohan @RohanShadman]

Current LLMs are hitting the ceiling on “more tokens = better thinking.”

A promising direction is procedural memory over ever-longer chains of thought—capturing recurring reasoning as reusable behaviors.

Think smarter, not just longer.
#AI #LLM #Reasoning #Efficiency #MLOps

[ShadmanRohan @RohanShadman]

New from Google Research✨: Learn Your Way🎒
Upload a 📚textbook/PDF → Interactive Lessons

🧭Mind maps ⚡Quizzes 🎧Audio lessons

📊11% better retention (78% vs. 67%) vs digital reader.

𝗧𝗵𝗼𝘂𝗴𝗵𝘁𝗳𝘂𝗹 𝗱𝗲𝘀𝗶𝗴𝗻 > 𝗺𝗼𝗿𝗲 𝘀𝗰𝗿𝗲𝗲𝗻 𝘁𝗶𝗺𝗲.

#AI #education #google

[ShadmanRohan @RohanShadman]

🎉 Just had an incredible experience attending The 63rd Annual Meeting of the Association for Computational Linguistics! 🎉 - via #Whova event app

[ShadmanRohan @RohanShadman]

@godofprompt Oh no,, you ss so,

[Matthew Berman @MatthewBerman]

1/ Google Research unveils new paper: "Titans: Learning to Memorize at Test Time"

It introduces human-like memory structures to overcome the limits of Transformers, with one "SURPRISING" feature.

Here's why this is huge for AI. 🧵👇

[Aaron Mueller @amuuueller]

What can mechanistic interpretability do for computational psycholinguists? @michaelwhanna and I took a stab at this question! We investigate garden path sentence processing in LMs at the feature (circuit) level.

[Wei Xu @cocoweixu]

We wrapped up CS 8803 "Large Language Model" class at @GeorgiaTech for Fall 2024.

Here is the reading list:

• learning from human preferences (PPO, DPO, SimPO, CPO, RRHF, ORPO, CTO)
• real-world LLM (Llama-3, Aya, Arena's)
• efficient LLM (MoMa, LoRA, QLoRA, LESS)

[ShadmanRohan @RohanShadman]

@MAarafat71 Star Jalsha shuru hoa gese😂

[Arpit Adlakha @arpit20adlakha]

One of the finest roadmaps I have seen for Senior Software Interviews, a guy posted on LeetCode for clearing Uber L5A, L5B or Google L5/L6 levels.

[Tim Denning @Tim_Denning]

I’ve spent over 120 hours studying one of the most controversial authors.

Nassim Taleb.

Here are 11 of his best lessons ↓

[Ole Lehmann @itsolelehmann]

I'm 32.

After living my whole life in Germany, last year I took the leap and moved abroad to Cyprus.

It's the greatest lifestyle upgrade I've ever experienced.

20 lessons for living the good life abroad (that'll make your move easier):

[ShadmanRohan @RohanShadman]

@rose_e_wang Will the workshop submissions be included in the conference proceedings, or are they considered non-archival?

[Toan Truong @ToanTruong_]

I'm 18.

I’m obsessed with learning how to learn.

So, I spent 200+ hours studying how geniuses, prodigies, and high performers master their disciplines.

Here's what I found on how to master anything faster: