Training BART Model on CPU instead of GPU

I am using the transformer’s trainer API to train a BART model on server. The GPU space is enough, however, the training process only runs on CPU instead of GPU.

I tried to use cuda and jit from numba like this example to add function decorators, but it still doesn’t help.

What is the reason of it using CPU instead of GPU? How can I solve it and make the process run on GPU?

Thank you for your help!