pytorch

babc862b - [functorch] Use nll_loss_backward decomposition for jvp transform (pytorch/functorch#764)