This is mathematically equivalent (the constant cancels) but ensures the largest exponent is $ e^0 = 1 $, preventing overflow. All major deep learning frameworks implement this stabilized version automatically.
