File size: 5,020 Bytes
64300ed
 
1d2dc7d
 
 
 
 
 
64300ed
1d2dc7d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7397626
d050b3f
1d2dc7d
 
 
 
 
 
6da2969
 
216d5f6
41e3a67
1d2dc7d
 
 
 
 
 
 
 
 
 
 
 
 
 
41e3a67
 
 
 
1d2dc7d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0226389
 
 
 
 
 
 
 
 
1d2dc7d
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
---
license: apache-2.0
language:
- en
- zh
pipeline_tag: text-generation
tags:
- ' TransNormerLLM'
---

<div align="center">
<h1>
  TransNormerLLM3 -- A Faster and Better LLM
</h1>
</div>

# Introduction

This official repository unveils the TransNormerLLM3 model along with its open-source weights for every 50 billion tokens processed during pre-training.

[TransNormerLLM](https://arxiv.org/abs/2307.14995) evolving from [TransNormer](https://arxiv.org/abs/2210.10340), standing out as the first LLM within the linear transformer architecture. Additionally, it distinguishes itself by being the first non-Transformer LLM to exceed both traditional Transformer and other efficient Transformer models (such as, RetNet and Mamba) in terms of speed and performance.


# TransNormerLLM3
- **TransNormerLLM3-15B** features **14.83 billion** parameters. It is structured with **42 layers**, includes **40 attention heads**, and has a total **embedding size of 5120**.
- **TransNormerLLM3-15B** is purely intergrated with **[Lightning Attention-2](http://arxiv.org/abs/2401.04658)**, which can maintain a **stable TGS** during training of **unlimited sequence lengths**, up until encountering firm limitations like GPU memory constraints.   
- **Titoken** tokenizer is used with a total **vocabulary size** of about **100,000**. 
- It incorporates **Simple GLU** for its channel mixer, **GLA** in the token mixer, and **SRMSNorm** for normalization.
- In terms of position encoding, the first layer employs **LRPE with exponential decay**, whereas the subsequent layers continue with **exponential decay encoding**.

### Pre-training Logbook
* Realtime Track: https://api.wandb.ai/links/opennlplab/kip314lq  
* Join to dicussion: [discord](https://discord.gg/MYQh6BWN) <<<>>> [wechat group](https://github.com/OpenNLPLab/TransnormerLLM/blob/main/images/contact_me_qr.png)
  
> --23.12.25-- startup: [WeChat - 预训练启航](https://mp.weixin.qq.com/s/YjUY-uy89WkF75_-rBTuKw)  <<<>>>  [Twitter - Pre-training Commences ](https://twitter.com/opennlplab/status/1739568669502611825) <<<>>> [YouTube Recording](https://t.co/wk7svS4o5r)   <<<>>> [bilibili 回放](https://www.bilibili.com/video/BV11j411J7Dy)  
> --24.01.02-- first week review: [WeChat - 第一周概览](https://mp.weixin.qq.com/s/zwGnZZI3itNPoxzzXkuU2w) <<<>>>   [Twitter - First Week Review](https://twitter.com/opennlplab/status/1742187694078501038)  
> --24.01.09-- second week review: [WeChat - 第二周概览](https://mp.weixin.qq.com/s/6D0qi-0aBier05OKuHfPEA) <<<>>>   [Twitter - Second Week Review](https://twitter.com/opennlplab/status/1744720007299523063)  


# Released Weights

|  param  | token | Hugging Face | Model Scope | Wisemodel |
| :-----: | :---: | :----------: | :---------: | :-------: |
| **15B** |  50B  |      🤗       |      🤖      |     🐯     |

# Benchmark Results
The evaluations of all models are conducted using the official settings and the [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) framework.

| Model                   | P   | T    | BoolQ | PIQA  | HS    | WG    | ARC-e | ARC-c | OBQA  | MMLU  | C-Eval |
| ----------------------- | --- | ---- | ----- | ----- | ----- | ----- | ----- | ----- | ----- | ----- | ------ |
| **TransNormerLLM3-15B** | 15  | 0.05 | 62.08 | 72.52 | 55.55 | 57.14 | 62.12 | 31.14 | 32.40 | 27.50 | 26.18  |
| **TransNormerLLM3-15B** | 15  | 0.10 | 63.98 | 74.70 | 61.09 | 61.33 | 65.95 | 34.64 | 35.60 | 25.38 | 27.40  |
| **TransNormerLLM3-15B** | 15  | 0.15 | 60.34 | 75.08 | 63.99 | 62.04 | 64.56 | 34.90 | 35.20 | 22.64 | 26.60  |



> **P**: parameter size (billion). **T**: tokens (trillion). **BoolQ**: acc. **PIQA**: acc. **HellaSwag**: acc_norm. **WinoGrande**: acc. **ARC-easy**: acc. **ARC-challenge**: acc_norm. **OpenBookQA**: acc_norm. **MMLU**: 5-shot acc. **C-Eval**: 5-shot acc.


# Acknowledgments and Citation

## Acknowledgments
Our project is developed based on the following open source projects:
- [tiktoken](https://github.com/openai/tiktoken) for the tokenizer.
- [metaseq](https://github.com/facebookresearch/metaseq) for training.
- [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) for evaluation.


## Citation
If you wish to cite our work, please use the following reference:
```
@article{qin2023scaling,
  title={Scaling transnormer to 175 billion parameters},
  author={Qin, Zhen and Li, Dong and Sun, Weigao and Sun, Weixuan and Shen, Xuyang and Han, Xiaodong and Wei, Yunshen and Lv, Baohong and Yuan, Fei and Luo, Xiao and others},
  journal={arXiv preprint arXiv:2307.14995},
  year={2023}
}

@misc{qin2024lightning,
      title={Lightning Attention-2: A Free Lunch for Handling Unlimited Sequence Lengths in Large Language Models}, 
      author={Zhen Qin and Weigao Sun and Dong Li and Xuyang Shen and Weixuan Sun and Yiran Zhong},
      year={2024},
      eprint={2401.04658},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```

<p align="center">
  - OpenNLPLab @2024 -
</p>