Each attention score is the dot product $q\cdot k$ of a query and a key with $d$ entries, and softmax turns the scores into positive weights that sum to one.

Before the softmax, every score is divided by $\sqrt{d}$, about 11.3 for $d=128$.