CUDA out of memory

Tried reducing the batch size to 4 but still…

What model are you using? It may be too large for 1 GPU. Try with batch size 1

I’m using XLM-RoBERTa-Large model for zero-shot classification.