In the masked Query-Key-Value (QKV) attention formula, a masking variable $\mathbf{Mask} \in \mathbb{R}^{(m+1) \times (m+1)}$ is added to the scaled dot-product of queries and keys before the softmax operation. This matrix ensures that each token only attends to itself and the tokens that precede it in the sequence. Specifically, to mask out future tokens, the entry $(i,j)$ of the mask corresponding to the query $\mathbf{q}_{i}$ and the key $\mathbf{k}_{j}$ is set to $-\infty$ if $i < j$.