}

  template <typename T>
  void TestResizedShapeForType(const OpType op_type, const bool align_corners,
                               const bool half_pixel_centers) {
    TensorShape x_shape({1, 2, 2, 1});
    Tensor x_data = MakeData<T>(x_shape);
    Output x, y;
    MakeOp<T>(op_type, x_data, {4, 6}, align_corners, half_pixel_centers, &x,
              &y);

    ClientSession session(scope_);
    std::vector<Tensor> outputs;

  ManipGradTest() : scope_(Scope::NewRootScope()) {}

  void RunTest(const Output& x, const TensorShape& x_shape, const Output& y,
               const TensorShape& y_shape) {
    TF_ASSERT_OK(scope_.status());
    float max_error;
    TF_ASSERT_OK((ComputeGradientError<float, float, float>(
        scope_, {x}, {x_shape}, {y}, {y_shape}, &max_error)));
    EXPECT_LT(max_error, 1e-4);
  }

  Scope scope_;
};

                            const std::vector<TensorShape>& x_shapes,
                            const OutputList& ys,
                            const std::vector<TensorShape>& y_shapes,
                            JAC_T* max_error) {
  if (xs.size() != x_shapes.size()) {
    return errors::InvalidArgument("xs(size ", xs.size(),
                                   ") and x_shapes(size ", x_shapes.size(),

    XlaLayoutPreference layout_preference) {
  xla::Shape xla_shape;
  TF_RETURN_IF_ERROR(
      tensorflow::TensorShapeToXLAShape(type, shape, &xla_shape));
  ApiConverter::StackHelper<XLA_Shape> se_shape(xla_shape);
  ApiConverter::StackHelper<XLA_Shape> tpu_shape;
  StatusHelper status;
  stream_executor::tpu::ExecutorApiFn()->XlaShapeToTpuShapeRepresentationFn(
      &se_shape.value, type, use_fast_memory, &tpu_shape.value,

          contracting_dims.add(c)
      x_signature = [
          None if c not in contracting_dims else x_shape[cidx]
          for cidx, c in enumerate(x_labels)
      ]
      y_signature = [
          None if c not in contracting_dims else y_shape[cidx]
          for cidx, c in enumerate(y_labels)
      ]
    return x_shape, y_shape, bias_shape, x_signature, y_signature

  def _create_einsum_model(
      self,

      x_signature = [
          None if c not in contracting_dims else x_shape[cidx]
          for cidx, c in enumerate(x_labels)
      ]
      y_signature = [
          None if c not in contracting_dims else y_shape[cidx]
          for cidx, c in enumerate(y_labels)
      ]
    return x_shape, y_shape, bias_shape, x_signature, y_signature

  def _create_einsum_model(
      self,

template <typename X_T, typename Y_T, typename JAC_T>
Status ComputeGradientError(const Scope& scope, const OutputList& xs,
                            const std::vector<TensorShape>& x_shapes,
                            const OutputList& ys,
                            const std::vector<TensorShape>& y_shapes,
                            JAC_T* max_error);

    x = Conj(scope, x);
    y = Conj(scope, y);
  }

  const auto x_shape = Shape(scope, x);
  const auto y_shape = Shape(scope, y);
  Output grad_x =
      EinsumGradWrt(scope, grad, y, x_shape, x_subs, y_subs, output_subs);
  Output grad_y =
      EinsumGradWrt(scope, grad, x, y_shape, y_subs, x_subs, output_subs);

  if (!absl::StrContains(output_subs, kEllipsis)) {

  // CHECK-DAG: %[[RHS:.+]] = mhlo.subtract %[[OFFSET]], %[[MUL_MEAN]] : tensor<?xf32>
  // CHECK-DAG: %[[X_SHAPE:.+]] = shape.shape_of %[[X]] : tensor<?x?x?x?xf32> -> tensor<4xindex>
  // CHECK-DAG: %[[MULTIPLIER_BCAST:.+]] = "mhlo.dynamic_broadcast_in_dim"(%[[MULTIPLIER]], %[[X_SHAPE]]) <{broadcast_dimensions = dense<1> : tensor<1xi64>}> : (tensor<?xf32>, tensor<4xindex>) -> tensor<?x?x?x?xf32>

  // The internal gradient implementation needs the shape of the input image.
  // x_shape = shape(x)[1:3]
  //         = slice(shape(x), {1}, {3 - 1})
  auto x_shape = Slice(scope, Shape(scope, op.input(0)), {1}, {2});
  grad_outputs->push_back(internal::ResizeNearestNeighborGrad(
      scope, grad_inputs[0], x_shape,
      internal::ResizeNearestNeighborGrad::AlignCorners(align_corners)