should both: attention_mask and global_attention_mask be used for classification?

Hi,

Again a conceptual question on text classification.

Since global attention is used on `````` only, I am slightly confused if I should just pass global_attention_mask to the model or both: attention_mask and global_attention_mask. I follow that attention_mask is mainly used to mask the `````` tokens but does it mean ``` n ``` ² complexity for local attention?

Thanks!