Where is the global attention?

[Trace2333]

I searched some code in longformer and the related code in transformers,the invert_mask() function in BartEncoder destroies the integer 2 in the attention mask,but the longformer attention code regard the mask as it dosen't been inverted.
So I think the global attention is not enabled in the model, could you explain it to me ?

Hope that I'm wrong....