cleverhans - tf2 implementation of momentum_iterative_method has an inconsistency with the algorithm from the paper

Question

tf2 implementation of momentum_iterative_method： line 93 of momentum_iterative_method.py use “tf.math.reduce_mean(tf.math.abs(grad), red_ind, keepdims=True)” for updating the grad.

but in the paper: Algorithm 1 uses L1 norm (reduce_sum()).

So, is this a mistake? or better using reduce_mean() here?