There is an issue with freezing parameters

[1190300611]

When I observed the parameter quantity and trainable parameter quantity of the model, I found that the trainable parameter quantity in the first stage was not normal

Extra parameters for tokenizer

The specific reason is that the anyToImageVideoAudio.py file first freezes the parameters of llama before initializing the tokenizer, which leads to the occurrence of this problem

[ChocoWu]

This is because we resize the embedding layer of LLM after adding new tokens, so the parameters of the embedding layer become trainable.