absl::Span<AbstractTensorHandle*> grad_inputs) override {
    // Grad for Softmax Input
    TF_RETURN_IF_ERROR(BroadcastMul(
        ctx, grad_outputs[0], forward_outputs_[1],
        grad_inputs.subspan(0, 1)));  // upstream_grad * local softmax grad

    // Grad for labels is null
    grad_inputs[1] = nullptr;
    return absl::OkStatus();
  }

Status AddWhileGradientLoop(WhileContext* while_ctx,
                            const std::vector<Output>& grad_inputs,
                            const Output& backprop_execution_pred,
                            const Scope& parent_scope,
                            std::vector<Output>* grad_outputs) {
  DCHECK_EQ(grad_inputs.size(), while_ctx->body_outputs().size());
  DCHECK_EQ(while_ctx->body_inputs().size(), while_ctx->body_outputs().size());