5fa1a76
1
2
3
4
This is shown in Figure 2d of the paper, see below for a sample attention mask: Using those attention matrices with less parameters then allows the model to have inputs having a bigger sequence length.
