/* Given upstream grad U and a mul op A*B, the gradients are:
     *
     *    dA = U * B
     *    dB = A * U
     *
     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];

    // Gradient for A
    std::string name = "Mul_Grad_A";
    TF_RETURN_IF_ERROR(Mul(ctx, upstream_grad, forward_inputs_[1],