It is trained from scratch on 2T English and Chinese tokens, and exhibits comparable performance with DeekSeek 7B and LLaMA2 7B, with only about 40% of computations.
For research purposes, we release the model checkpoints of DeepSeekMoE 16B Base and DeepSeekMoE 16B Chat to the public, which can be deployed on a single GPU with 40GB of memory without the need for quantization.
The model code file can be found [here](https://huggingface.co/deepseek-ai/deepseek-moe-16b-base/blob/main/modeling_deepseek.py).
## 2. Evaluation Results
### DeepSeekMoE 16B Base
We evaluate DeepSeekMoE 16B on various benchmarks and compare it with a series of models, as shown in the following.
- Comparison with open source models on the Open LLM Leaderboard. DeepSeekMoE 16B consistently outperforms models with a similar number of activated parameters by a large margin, and achieves comparable performance with LLaMA2 7B, which has approximately 2.5 times the activated parameters.
- Comparison with DeepSeek 7B on our internal benchmarks. DeepSeek 7B is a dense model trained on the same corpus as DeepSeekMoE 16B. With only 40.5% of computations, DeepSeekMoE 16B achieves comparable performance with DeepSeek 7B.
- Comparison with LLaMA2 7B on our internal benchmarks. With only 39.6% of computations, DeepSeekMoE 16B outperforms LLaMA2 7B on the majority of benchmarks.
We also evaluate DeepSeekMoE 16B Chat on various benchmarks and compare it with DeepSeek 7B Chat and LLaMA2 7B SFT. All of the compared models follow the same fine-tuning setting and data for fair comparison.
The evaluation results are shown in the following. With only about 40% of computations, DeepSeekMoE 16B Chat achieves comparable or better performance than DeepSeek 7B Chat and LLaMA2 7B SFT.
We release the DeepSeekMoE 16B, including both base and chat models, to the public. To support a broader and more diverse range of research within both academic and commercial communities. Please **note** that the use of this model is subject to the terms outlined in [License section](#5-license). Commercial usage is permitted under these terms.
text = "An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is"
result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True)
print(result)
```
Avoiding the use of the provided function `apply_chat_template`, you can also interact with our model following the sample template. Note that `messages` should be replaced by your input.
**Note:** By default (`add_special_tokens=True`), our tokenizer automatically adds a `bos_token` (`<｜begin▁of▁sentence｜>`) before the input text. Additionally, since the system prompt is not compatible with this version of our models, we DO NOT RECOMMEND including the system prompt in your input.
### How to Fine-tune DeepSeekMoE
We provide script `fintune/finetune.py` for users to finetune our models on downstream tasks.
The script supports the training with [DeepSpeed](https://github.com/microsoft/DeepSpeed). You need install required packages by:
```bash
pip install -r requirements.txt
```
Please follow [Sample Dataset Format](https://huggingface.co/datasets/garage-bAInd/Open-Platypus) to prepare your training data.
Each item has two required fields `instruction` and `output`.
After data preparation, you can use the sample shell script to finetune the DeepSeekMoE model.
Remember to specify `DATA_PATH`, `OUTPUT_PATH`.
And please choose appropriate hyper-parameters(e.g., `learning_rate`, `per_device_train_batch_size`) according to your scenario.
We have used flash_attention2 by default. For devices supported by flash_attention, you can refer [here](https://github.com/Dao-AILab/flash-attention).
For this configuration, zero_stage needs to be set to 3, and we run it on eight A100 40 GPUs.
You can also finetune the model with 4/8-bits qlora, feel free to try it. For this configuration, it is possible to run on a single A100 80G GPU, and adjustments can be made according to your resources.
This code repository is licensed under the MIT License. The use of DeepSeekMoE models is subject to the Model License. DeepSeekMoE supports commercial use.
See the [LICENSE-CODE](LICENSE-CODE) and [LICENSE-MODEL](LICENSE-MODEL) for more details.
## 6. Citation
```
@article{dai2024deepseekmoe,
author={Damai Dai and Chengqi Deng and Chenggang Zhao and R. X. Xu and Huazuo Gao and Deli Chen and Jiashi Li and Wangding Zeng and Xingkai Yu and Y. Wu and Zhenda Xie and Y. K. Li and Panpan Huang and Fuli Luo and Chong Ruan and Zhifang Sui and Wenfeng Liang},
title={DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models},
journal = {CoRR},
volume = {abs/2401.06066},
year = {2024},
url = {https://arxiv.org/abs/2401.06066},
}
```
## 7. Contact
If you have any questions, please raise an issue or contact us at [service@deepseek.com](mailto:service@deepseek.com).
