huggingface model how to convert to SAT model
Feature request / 功能建议
how to convert huggingface model convert to SAT model?
Motivation / 动机
So can use sat to train V2 faster.
Your contribution / 您的贡献
NA
Here are the SAT models of CogVLM2: https://github.com/THUDM/CogVLM/blob/cogvlm2_dev/utils/models/cogvlm2_model.py
You can train the model with this branch: https://github.com/THUDM/CogVLM/tree/cogvlm2_dev
Great,thank you so much!
Do I just need to modify this file finetune_demo/finetune_cogvlm_demo.py from FineTuneTrainCogVLMModel to FineTuneTrainCogVLM2Model? And I wonder how to intial the tokenizer?
Here are the SAT models of CogVLM2: https://github.com/THUDM/CogVLM/blob/cogvlm2_dev/utils/models/cogvlm2_model.py
You can train the model with this branch: https://github.com/THUDM/CogVLM/tree/cogvlm2_dev
It seems can't load directly current open-source CogVLM2 model in CogVLM finetune code, which still need to convert hf checkpoint to sat format for using.
请问您解决了吗,我下载不下来cogvlm2-llama3-chinese-chat-19Bzip文件,也希望能转一下