// And Args must have `index` attribute.
      TF_RETURN_IF_ERROR(GetNodeAttr(n->attrs(), "index", &index));
    } else if (n->type_string() == "Placeholder") {
      // Use custom attribute (prefixed with `_`) `_index` for placeholders as
      // they come from user specifications.
      if (const auto s = GetNodeAttr(n->attrs(), "_index", &index); !s.ok()) {

}

absl::flat_hash_map<string, string> GetXlaInternalScopes(const Graph& graph) {
  absl::flat_hash_map<string, string> scopes;
  for (Node* node : graph.nodes()) {
    string scope;
    if (GetNodeAttr(node->attrs(), kXlaInternalScopeAttr, &scope).ok()) {
      scopes[node->name()] = scope;
    }
  }

  if (VLOG_IS_ON(2)) {
    VLOG(2) << "_XlaInternalScopes:";
    for (const auto& p : scopes) {

    return false;
  }
  return guaranteed_constant;
}

// Finds the `index` of an _Arg or _Retval node.
Status GetIndexAttr(const Node& n, int num_args, int* index) {
  TF_RETURN_IF_ERROR(GetNodeAttr(n.attrs(), "index", index));
  if (*index < 0 || *index >= num_args) {
    return errors::InvalidArgument("Invalid ", n.type_string(), " number ",
                                   *index);
  }
  return absl::OkStatus();

  // The desired propagation of the gradients back to the data inputs is:
  // [g1, g2, g4, g5, g3]
  auto data = op.input(0);
  auto partitions = op.input(1);
  int32_t num_partitions;
  TF_RETURN_IF_ERROR(
      GetNodeAttr(op.node()->attrs(), "num_partitions", &num_partitions));

  // Note: the shape of the partitions is a prefix of the data shape.
  // shape(partitions) = [5]
  auto partitions_shape = Shape(scope, partitions);

 private:
  Graph* graph_;
  OptimizerOptions::GlobalJitLevel global_jit_level_;
  size_t unique_scope_id_;
};

std::optional<string> GetXlaInternalScope(Node* node) {
  string scope;
  if (GetNodeAttr(node->attrs(), kXlaInternalScopeAttr, &scope).ok()) {
    return scope;
  }

  return std::nullopt;
}

void SetXlaInternalScope(Node* node, StringPiece scope) {
  node->AddAttr(kXlaInternalScopeAttr, scope);

  }

  new_in->set_assigned_device_name(n->assigned_device_name());
  return new_in;
}

namespace {
absl::StatusOr<bool> IsConstantSmall(Node* n) {
  const TensorProto* proto = nullptr;
  TF_RETURN_IF_ERROR(GetNodeAttr(n->def(), "value", &proto));

  int64_t total_elements = 1;
  for (const auto& dim : proto->tensor_shape().dim()) {
    if (dim.size() < 0) {
      return errors::Internal("Unknown dimension size in constant tensor ",

  return BiasAdd(cop_scopes.last, m, b);
}

void GetColocationConstraints(const Output& tensor,
                              std::vector<string>* constraints) {
  constraints->clear();
  TF_EXPECT_OK(GetNodeAttr(tensor.op().node()->attrs(), kColocationAttrName,
                           constraints));
}

TEST(CCOpTest, Basic) {
  Scope root = Scope::NewRootScope();
  auto c = Const(root, {{1, 1}});

    a = ConjugateHelper(scope, a);
    b = ConjugateHelper(scope, b);
  }
  auto product = op.output(0);

  bool ta;
  bool tb;
  TF_RETURN_IF_ERROR(GetNodeAttr(product.node()->attrs(), attr_adj_x, &ta));
  TF_RETURN_IF_ERROR(GetNodeAttr(product.node()->attrs(), attr_adj_y, &tb));

  if (!ta && !tb) {
    return MatMulGradHelper(scope, is_batch, grad_inputs[0], false, b, true,

}

Status MakeCallNodeFromAttribute(const Node& node, const std::string& attr_name,
                                 NodeDef* node_def) {
  const NameAttrList* name_attr;
  TF_RETURN_IF_ERROR(GetNodeAttr(node.attrs(), attr_name, &name_attr));
  node_def->set_op(name_attr->name());
  *(node_def->mutable_attr()) = name_attr->attr();
  return absl::OkStatus();
}

StatusOrOptional<Tensor> TryToGetTensorFromConstOp(Node* n) {
  if (n->type_string() != "Const") {
    return {std::nullopt};
  }

  const TensorProto* proto = nullptr;
  TF_RETURN_IF_ERROR(GetNodeAttr(n->def(), "value", &proto));
  Tensor tensor(proto->dtype());
  TF_RET_CHECK(tensor.FromProto(*proto));
  return {tensor};
}

struct SliceInputs {
  Output slice_op;
  Output input;
  Output begin;