📚 Documentation

The attention mask shape should be (L, S) since it is added to the attention weights which is of shape (N, L, S). This is only in master, and not on pytorch.org yet.

#20850