Is it possible to fully fine tune the model on a GPU of 32GB by reducing batch size only?

[frandoLin]

I noticed that it is required to use at 70GB to fully fine tune the model with a batch size of 16, so I am wondering if it is possiable to get it work on a 32GB GPU by reducing the batch size only. If no, are there any suggestions to achieve the goal? Thanks!

[ModiShi]

You can try that, and if you still run out of memory, I recommend fine-tuning only the action expert.

[frandoLin]

You can try that, and if you still run out of memory, I recommend fine-tuning only the action expert.

Okay, but do you say the model performance is roughly the same after fine-tuning the action expert only?

[ModiShi]

Fine-tuning the full model would be better.