How to increase the batch size of the pretrained mBART50 model?

I have created a database of 100k tables and I want to translate them using mBART50. I have access to 16GB GPU but currently, the model is only using 4GB, hence how to increase the batch size?