It actually returns sub-gradients because indeed it is not differentiable but similarly to relu that is not differentiable, sub-gradient are quite sufficient for many optimization tasks (but less stable especially if you compute gradients wrt weights)