*    dA = U * B.T
     *    dB = A.T * U
     *
     *    where A.T means `transpose(A)`
     */
    AbstractTensorHandle* upstream_grad = grad_outputs[0];

    // Get transpose attrs
    bool t_a;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("transpose_a", &t_a));

    bool t_b;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("transpose_b", &t_b));

    // Conj each input
    AbstractTensorHandle* conj_output;

                         {6, 15, 24, 6, 15, 24, 6, 15, 24}};

  for (int i{}; i < 4; ++i) {
    bool transpose_a = transpose_a_vals[i];
    bool transpose_b = transpose_b_vals[i];
    Model MatMulModel =
        [transpose_a, transpose_b](
            AbstractContext* ctx,
            absl::Span<AbstractTensorHandle* const> inputs,

TF_CAPI_EXPORT extern int TFE_TensorDebugInfoOnDeviceNumDims(
    TFE_TensorDebugInfo* debug_info);

// Returns the number of elements in dimension `dim_index`.
// Tensor representation on device can be transposed from its representation
// on host. The data contained in dimension `dim_index` on device
// can correspond to the data contained in another dimension in on-host