value = -1;

  // Read our first value
  cache2.LookupOrCompute("a", &value, compute_func, &status);
  TF_EXPECT_OK(status.status);
  EXPECT_EQ(value, 0);
  EXPECT_EQ(num_compute_calls, 1);
  // Re-read, exepct no additional function compute_func calls.
  cache2.LookupOrCompute("a", &value, compute_func, &status);
  TF_EXPECT_OK(status.status);
  EXPECT_EQ(value, 0);
  EXPECT_EQ(num_compute_calls, 1);

    func_args.push_back(TF_CreateAbstractTensorFromEagerTensor(input_eager, s));
    ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
  }

  TF_OutputList* func_outputs = TF_NewOutputList();
  TF_OutputListSetNumOutputs(func_outputs, 2, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
  TF_ExecuteOperation(fn_op, func_args.size(), func_args.data(), func_outputs,
                      s);

 public:
  explicit CGraphOptimizer(TP_Optimizer optimizer, const char* device_type)
      : optimizer_(optimizer), device_type_(device_type) {
    if (optimizer.create_func != nullptr) {
      c_optimizer_ = (*optimizer_.create_func)();
    } else {
      c_optimizer_ = nullptr;
    }
  }
  std::string name() const override { return "PluggableGraphOptimizer"; }

// is not ready when we start running an op or a function.
void TestRemoteExecuteSilentCopies(bool async, bool remote, bool func,
                                   bool heavy_load_on_streaming_rpc,
                                   bool remote_func_outputs = false,
                                   bool has_packed_input = false);

void TF_DeleteAbstractFunction(TF_AbstractFunction* func) {
  unwrap(func)->Unref();
}

void TF_ExecutionContextRegisterFunction(TF_ExecutionContext* ctx,
                                         TF_AbstractFunction* func,
                                         TF_Status* s) {
  tsl::Set_TF_Status_from_Status(s,
                                 unwrap(ctx)->RegisterFunction(unwrap(func)));

  absl::StatusOr<core::RefCountPtr<FunctionRecord>> GetFunctionRecord()
      override {
    return func_record_.GetNewRef();
  }

  // For LLVM style RTTI.
  static bool classof(const AbstractFunction* ptr) {
    return ptr->getKind() == kGraph;
  }

 private:
  core::RefCountPtr<FunctionRecord> func_record_;
};
}  // namespace graph
}  // namespace tracing
}  // namespace tensorflow

// destructor. The easiest way to use MakeCleanup is with a lambda argument,
// capturing the return value in an 'auto' local variable. Most users will not
// need more sophisticated syntax than that.
//
// Example:
//   void func() {
//     FILE* fp = fopen("data.txt", "r");
//     if (fp == nullptr) return;
//     auto fp_cleaner = gtl::MakeCleanup([fp] { fclose(fp); });
//     // No matter what, fclose(fp) will happen.
//     DataObject d;

}

void TF_SetAttrFuncName(TF_OperationDescription* desc, const char* attr_name,
                        const char* value, size_t length) {
  tensorflow::NameAttrList func_name;
  func_name.set_name(string(value, value + length));
  desc->node_builder.Attr(attr_name, func_name);
}

void TF_SetAttrShape(TF_OperationDescription* desc, const char* attr_name,
                     const int64_t* dims, int num_dims) {

  }
  const auto len = std::min(max_func, static_cast<int>(lib.function_size()));
  for (int i = 0; i < len; ++i) {
    TF_Function* func = new TF_Function();
    func->record = new tensorflow::FunctionRecord(lib.function(i), {}, false);
    funcs[i] = func;
  }
  status->status = absl::OkStatus();
  return len;
}

void TF_FunctionToFunctionDef(TF_Function* func, TF_Buffer* output_func_def,

  params->optimizer_configs->struct_size = TP_OPTIMIZER_CONFIGS_STRUCT_SIZE;
  params->optimizer->struct_size = TP_OPTIMIZER_STRUCT_SIZE;
  params->optimizer->create_func = nullptr;
  params->optimizer->optimize_func = optimize_func;
  params->optimizer->destroy_func = nullptr;
}

TEST(Grappler, SuccessfulRegistration) {
  auto plugin_init = [](TP_OptimizerRegistrationParams* const params,