Update README.md
Browse files
README.md
CHANGED
|
@@ -1,58 +1,62 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: mit
|
| 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 |
```
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
pipeline_tag: image-text-to-text
|
| 4 |
+
tags:
|
| 5 |
+
- text-generation-inference
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
<h2 align="center"> <a href="https://arxiv.org/abs/2405.14297">Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models</a></h2>
|
| 9 |
+
<h5 align="center"> If our project helps you, please give us a star ⭐ on <a href="https://github.com/LINs-lab/DynMoE">GitHub</a> and cite our paper!</h2>
|
| 10 |
+
<h5 align="center">
|
| 11 |
+
|
| 12 |
+
## 📰 News
|
| 13 |
+
|
| 14 |
+
- **[2024.5.31]** 🔥 Our [code](https://github.com/LINs-lab/DynMoE/) is released!
|
| 15 |
+
- **[2024.05.25]** 🔥 Our **checkpoints** are available now!
|
| 16 |
+
- **[2024.05.23]** 🔥 Our [paper](https://arxiv.org/abs/2405.14297) is released!
|
| 17 |
+
|
| 18 |
+
## 😎 What's Interesting?
|
| 19 |
+
|
| 20 |
+
**Dynamic Mixture of Experts (DynMoE)** incorporates (1) a novel gating method that enables each token to automatically determine the number of experts to activate. (2) An adaptive process automatically adjusts the number of experts during training.
|
| 21 |
+
|
| 22 |
+
### Top-Any Gating
|
| 23 |
+
|
| 24 |
+
<video controls src="https://i.imgur.com/bLgNaoH.mp4" title="Top-Any Gating"></video>
|
| 25 |
+
|
| 26 |
+
### Adaptive Training Process
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
## 💡 Model Details
|
| 31 |
+
|
| 32 |
+
- 🤔 DynMoE-Qwen is a MoE model with **dynamic top-k gating**, finetuned on [LanguageBind/MoE-LLaVA-Qwen-Stage2](https://huggingface.co/LanguageBind/MoE-LLaVA-Qwen-Stage2).
|
| 33 |
+
- 🚀 Our DynMoE-Qwen-1.8B has totally 3.1B parameters, but **only 2.2B are activated!** (average top-k = 1.86)
|
| 34 |
+
- ⌛ With the DynMoE tuning stage, we can complete training on 8 A100 GPUs **within 40 hours.**
|
| 35 |
+
|
| 36 |
+
## 👍 Acknowledgement
|
| 37 |
+
|
| 38 |
+
We are grateful for the following awesome projects:
|
| 39 |
+
|
| 40 |
+
- [tutel](https://github.com/microsoft/tutel)
|
| 41 |
+
- [DeepSpeed](https://github.com/microsoft/DeepSpeed)
|
| 42 |
+
- [GMoE](https://github.com/Luodian/Generalizable-Mixture-of-Experts)
|
| 43 |
+
- [EMoE](https://github.com/qiuzh20/EMoE)
|
| 44 |
+
- [MoE-LLaVA](https://github.com/PKU-YuanGroup/MoE-LLaVA)
|
| 45 |
+
- [GLUE-X](https://github.com/YangLinyi/GLUE-X)
|
| 46 |
+
|
| 47 |
+
## 🔒 License
|
| 48 |
+
|
| 49 |
+
This project is released under the MIT license as found in the [LICENSE](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/mit.md) file.
|
| 50 |
+
|
| 51 |
+
## ✏️ Citation
|
| 52 |
+
|
| 53 |
+
```tex
|
| 54 |
+
@misc{guo2024dynamic,
|
| 55 |
+
title={Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models},
|
| 56 |
+
author={Yongxin Guo and Zhenglin Cheng and Xiaoying Tang and Tao Lin},
|
| 57 |
+
year={2024},
|
| 58 |
+
eprint={2405.14297},
|
| 59 |
+
archivePrefix={arXiv},
|
| 60 |
+
primaryClass={cs.LG}
|
| 61 |
+
}
|
| 62 |
```
|