#if !defined(IS_MOBILE_PLATFORM) && !defined(IS_SLIM_BUILD)
  delete server;
#endif  // !defined(IS_MOBILE_PLATFORM) && !defined(IS_SLIM_BUILD)
}

void TF_RegisterLogListener(void (*listener)(const char*)) {
#if !defined(IS_MOBILE_PLATFORM) && !defined(IS_SLIM_BUILD)
  tensorflow::logging::RegisterListener(listener);
#endif  // !defined(IS_MOBILE_PLATFORM) && !defined(IS_SLIM_BUILD)
}

                                          const char* name) {
  const auto& m = reader->GetVariableToShapeMap();
  return m.at(name).dims();
}

// This builder is used in the eager API to build a NodeDef.
struct TF_AttrBuilder : public tensorflow::AttrBuilder {
  using tensorflow::AttrBuilder::AttrBuilder;
  // The string buffers to make sure that any `attr_name` we pass into

  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  // Build an abstract operation.
  auto* op = TF_NewAbstractOp(ctx);
  TF_AbstractOpSetOpType(op, "Add", status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  // Build inputs and outputs.
  TF_AbstractTensor* inputs[2] = {at, at};
  TF_OutputList* o = TF_NewOutputList();

// TODO(ashankar): Use gmock matchers instead?
// (https://github.com/google/googletest/blob/master/googlemock/docs/CookBook.md#writing-new-parameterized-matchers-quickly)
// That will require setting up the tensorflow build with gmock.
#define EXPECT_TF_META(attr_name, expected_list_size, expected_type, \
                       expected_total_size)                          \
  do {                                                               \

  // When we turn on grappler, confirm that the tf.Add has been rewritten into a
  // tf.Mul.
  // This capability of checking the executed op names is currently only enabled
  // for TFRT debug build, for performance and simplicity reasons.
  if (use_tfrt) {
    TF_Buffer* buf = TF_NewBuffer();
    TFE_GetExecutedOpNames(ctx, buf, status);
    ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

};

TEST_P(CppGradients, TestIdentityNGrad) {
  // This test is interesting because the current implementation of GradientTape
  // would return [0, 1] whereas we use build_default_zeros_grads=false here
  // so we get back [nullptr, 1].

  AbstractTensorHandlePtr x1;
  {
    AbstractTensorHandle* x1_raw = nullptr;
    status_ = TestScalarTensorHandle<float, TF_FLOAT>(

    }
  }

  TF_RETURN_IF_ERROR(GradientTape::ComputeGradient(
      vspace, target_tensor_ids, source_tensor_ids, sources_that_are_targets,
      output_gradients, result, /*build_default_zeros_grads*/ false));
  return absl::OkStatus();
}

// Helper functions which delegate to `AbstractOperation`, update
// the state of the ForwardOperation and call the tape as appropriate.

 private:
  friend class GraphContext;  // For access to op_.
  TF_Graph* g_;
  std::unique_ptr<TF_OperationDescription> op_;
  // Hold `op_type` and `op_name` till both are available since we need both
  // to build a graph operation.
  string op_type_;
  const char* op_name_ = nullptr;
  // TODO(srbs): Use this.
  string device_name_;
};

// GraphContext wraps a TF_Graph modeling a single function and manages the

      const tensorflow::FunctionDef& fdef,
      const std::vector<EdgeSpec>& e_edges,  // expected edges
      const std::vector<EdgeSpec>& c_edges,  // expected ctrl edges
      bool is_exact_edges = true) {
    // Build a set of edges from fdef
    std::set<EdgeSpec> a_edges;  // actual edges
    // Get edges from inputs to body nodes and between body nodes
    for (const NodeDef& node_def : fdef.node_def()) {