std::vector<Output>* grad_outputs) {
  // AddN doesn't support broadcasting, so all the inputs must be the
  // same shape.
  // Note:
  // dy/dx_k = d(x_1 + x_2 + ... + x_n)/dx_k = 1 for all x_k
  // hence dx_k = dy for all x_k
  // So the gradient for AddN just transfers the incoming gradient to
  // all outgoing gradients.
  auto incoming = Identity(scope, grad_inputs[0]);