Danyang Liu

Multimodal Agents, NLP | Member of Technical Staff @ Reka AI | PhD student @EdinburghNLP; Opinions my own.

[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

[will brown @willcb]

https://t.co/gCIFKAjB0Z

[Danyang Liu @dannyleoo]

@smdvln @MSFTResearch When is the expected start date for this position?

[Yulong Chen @Yulongchen1010]

Evaluating LLMs usually requires sophisticated human designs and with the continuous improvement of LLMs, it is difficult for humans to find their limitations.

Can LLMs find their own limitations by proposing questions to themselves? Check our new paper: https://t.co/lDkBCQsIVR

[Danyang Liu @dannyleoo]

@Yulongchen1010 Great work Yulong! One button three combo! 🔥

[Danyang Liu @dannyleoo]

From my experiments, they can hardly generalize to tasks not included in the instruction tuning dataset, meaning they are highly biased towards captioning/OCR/VQA tasks.

[Danyang Liu @dannyleoo]

Does instruction tuning of large vision language models really make sense? My understanding of instruction tuning is that it can allow the model to exhibit strong generalization capabilities, but the LLaVA family and other models don't seem to be good zero-shot learners.

[Jianhao (Elliott) Yan @yan_elliott]

Can GPT-4 serve as expert translators? No, Not Yet a Match for Expert Translators

In our latest preprint, we benchmark GPT-4 against human translators at various levels of expertise and professions.
https://t.co/4RIge7hUMe

[David Stutz @davidstutz92]

By popular request, I wrote a separate article on effectively managing and running experiments during a PhD in AI. I learned that effectively running experiments can be a crucial aspect of a PhD and I wanted to share some of my learnings 🧵: https://t.co/zDN4MAxlbC

[Danyang Liu @dannyleoo]

@NielsRogge @victorsanh Thank you for your response! It would be very helpful if there is a notebook for multi-GPU training and loading for inference. BTW, do you have any suggestions for freezing more layers, like which ones to prioritize?

[Ehud Reiter @EhudReiter]

New blog: Can LLM-based eval replace human evaluation?
https://t.co/rofFzdwSRh

[Danyang Liu @dannyleoo]

@vlachos_nlp @ZhenyunDeng @MoyYuan @pietro_lesci @Yulongchen1010 @cambridgenlp @FEVERworkshop Hi Andreas, when the expected start date is?

[Danyang Liu @dannyleoo]

@mervenoyann @llamafactory_ai yeah you can see here a summary

[Danyang Liu @dannyleoo]

@mervenoyann @llamafactory_ai I tried repeating the image in the input twice, and the peak GPU usage increased to ~30GB under the qlora setting (5 images -> 60GB). The model only takes 7GB. Does that mean adding one image to the input will use an additional ~3GB of memory? do_splitting is set to False.

[Danyang Liu @dannyleoo]

@mervenoyann @llamafactory_ai Yes, I'm using the tutorial you released: https://t.co/tP49WSJUZq

[Danyang Liu @dannyleoo]

@mervenoyann @llamafactory_ai Thanks for the reply! I tried fine-tuning idefics2 using the colab tutorial, but I noticed some unusual GPU memory usage. With qlora and bsz 2, a single image input takes 17GB, but five images input use around 61GB. Is this normal? I'd appreciate your thoughts on this!

[Sylvain Filoni @fffiloni]

✨Consistent Character @gradio demo portage is available on @huggingface

(adapted from the original cog by @fofrAI 👋)

[Xiao-Xiao Long @xxlong0]

Create your 3D portrait with only a single image using our powerful model Era3D. @ylecun

Era3D is a new high-resolution cross-domain diffusion model, which produces 512x512 color images and normal maps (Wonder3D and Syncdreamer only support 256x256 resolution). With such 2D representations, it's feasible to create such detailed 3D reconstruction. Many thanks to the efforts from @pengli_hkust

The reconstruction part cannot be integrated into huggingfaca demo. Vist our github repo for complete experiences.

Github repo: https://t.co/npB34quAIe
Webpage: https://t.co/3cDxi3GVri
Huggingface demo for MV color and normal prediction : https://t.co/zEt9KlAgYM