Implementations usually replace replace the 1 in the denominator with exp(-max(x)) for this reason.