torch:div

(torch:div a b)

Differentiable elementwise a / b with numpy-style broadcasting (linalg:div): the numerator's gradient is g / b, the denominator's -g * a / b^2. An array divided by a plain number (or an untracked scalar tensor) is returned as a view: nothing is computed until something reads the data, and a torch:softmax over it folds the division into its own pass.