why the gradient become very few after neuralsort ?

[buptzzl]

I try to do SGD optimizer after neuralsort operation, but after this operation, the gradient backup to dense layer become very few, like for 0.23 to -1.12376988e-07 . why ? how to enlarge the gradient.
Thanks very much!