Emova-ollm/deepseek-vl2-deepseekmoe-tiny_add_speech_token_4096_nostrip

DeepSeek-VL2-DeepSeekMoE-Tiny-Add-Speech-Token-4096-Nostrip

Introduction

This repo contains the DeepSeek-VL2-DeepSeekMoE-Tiny-Add-Speech-Token-4096-Nostrip model utilized to train the EMOVA series of models. Different from traditional LLMs based on dense Transformers, DeepSeekMoE LLMs utilize an efficient sparse Mixture-of-Experts (MoE) architecture. In total, it contains 3B parameters, while only a 0.57B subset is activated for each token during inference. This checkpoint is extracted from the DeepSeek-VL2-Tiny model.

Based on the original DeepSeek-VL2-DeepSeekMoE-Tiny checkpoint, we further insert speech tokens into its vocabulary for end-to-end omni-modal alignment as follows. The total number of speech tokens in EMOVA speech tokenizer is 4096. Thus, it should be utilized as initialization in the Stage 2: Omni-modal text-centric alignment of EMOVA training.

# Source code can be found https://github.com/emova-ollm/EMOVA#insert-speech-tokens-into-llm-vocabulary
python scripts/insert_speech_token.py \
  --origin_model_path Emova-ollm/deepseek-vl2-deepseekmoe-tiny \
  --saved_model_path ./deepseek-vl2-deepseekmoe-tiny_add_speech_token_4096_nostrip \
  --num_speech_tokens 4096

Usage

To train EMOVA with DeepSeek-VL2-DeepSeekMoE-Tiny-Add-Speech-Token-4096-Nostrip, we need to create a new model config, and set the language_model parameters as follows. An example is provided here. Check more details on training EMOVA in our github repo.

language_model=dict(
  type='EmovaDeepseekV2ForCausalLM',                                         -- Wrapper class type for EMOVA
  pretrained_model_name_or_path='Emova-ollm/deepseek-vl2-deepseekmoe-tiny_add_speech_token_4096_nostrip',  -- HuggingFace repo of pre-trained LLM
  attn_implementation="flash_attention_2",                                   -- Attention type
  from_pretrained=True,                                                      -- Load pre-trained weights
),

Citation

@article{chen2024emova,
  title={Emova: Empowering language models to see, hear and speak with vivid emotions},
  author={Chen, Kai and Gou, Yunhao and Huang, Runhui and Liu, Zhili and Tan, Daxin and Xu, Jing and Wang, Chunwei and Zhu, Yi and Zeng, Yihan and Yang, Kuo and others},
  journal={arXiv preprint arXiv:2409.18042},
  year={2024}
}

@article{wu2024deepseek,
  title={Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding},
  author={Wu, Zhiyu and Chen, Xiaokang and Pan, Zizheng and Liu, Xingchao and Liu, Wen and Dai, Damai and Gao, Huazuo and Ma, Yiyang and Wu, Chengyue and Wang, Bingxuan and others},
  journal={arXiv preprint arXiv:2412.10302},
  year={2024}
}