return;
    }

    RunMetadata run_metadata_proto;
    result = session->Run(run_options_proto, input_pairs, output_tensor_names,
                          target_oper_names, &outputs, &run_metadata_proto);

    // Serialize back to upstream client, who now owns the new buffer
    if (run_metadata != nullptr) {
      status->status = MessageToBuffer(run_metadata_proto, run_metadata);

  std::unique_ptr<ParallelTensor> run_collective =
      parallel_device.ScalarsFromSequence<bool>({true, true}, context.get(),
                                                status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  auto outputs = parallel_device.Execute(
      context.get(), {reduced_values.get(), run_collective.get()},

  TF_Tensor* ret;
  TF_SessionRun(session, /*run_options*/ nullptr,
                // input related parameters
                /*inputs*/ nullptr, /*input_values*/ nullptr, /*ninputs*/ 0,
                // output related parameters
                /*outputs*/ &output, /*output_values*/ &ret,
                /*noutputs*/ 1,
                /*targets*/ nullptr, /*ntargets*/ 0,
                /*run_metadata*/ nullptr, status);

  targets_.clear();
  for (TF_Operation* t : targets) {
    targets_.emplace_back(t);
  }
}

void CSession::Run(TF_Status* s) {
  if (inputs_.size() != input_values_.size()) {
    ADD_FAILURE() << "Call SetInputs() before Run()";
    return;
  }
  ResetOutputValues();
  output_values_.resize(outputs_.size(), nullptr);

  TF_Buffer* run_options = TF_NewBufferFromString("", 0);
  TF_Buffer* run_metadata = TF_NewBuffer();
  TF_Run(session, run_options, nullptr, nullptr, 0, nullptr, nullptr, 0,
         nullptr, 0, run_metadata, s);
  EXPECT_EQ(TF_INVALID_ARGUMENT, TF_GetCode(s)) << TF_Message(s);
  EXPECT_EQ("Session was not created with a graph before Run()!",
            string(TF_Message(s)));

    TF_DeleteGraph(host_graph_);
    TF_DeleteGraph(func_graph_);
    TF_DeleteStatus(s_);
  }

  void Run(const std::vector<std::pair<TF_Operation*, TF_Tensor*>>& inputs,
           TF_Operation* output, int32_t expected_result) {
    Run(inputs, {{output, 0}}, {expected_result});
  }

  // Run the host graph, which now contains a function and check that
  // outputs are as expected.

        op_(nullptr),
        thread_(tensorflow::Env::Default()->StartThread(
            tensorflow::ThreadOptions(), "parallel_device_execute",
            std::bind(&DeviceThread::Run, this))) {}
  ~DeviceThread();

  // Requests that the worker thread execute the specified operation. Blocks
  // until the previously pending operation (a StartExecute without a Join) has
  // finished, if any.
  //

            composite_device_name);
  EXPECT_EQ(TFE_TensorHandleBackingDeviceName(packed_handle, status),
            composite_device_name);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

  // Register and run a function which returns the sum of 3 variables.
  const string function_def = AddVariablesFunction();
  TFE_ContextAddFunctionDef(ctx, function_def.data(), function_def.size(),
                            status);

  TFE_DeleteContextOptions(opts);

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

  // Build the abstract op to run the function.
  TF_AbstractOp* fn_op = TF_NewAbstractOp(eager_execution_ctx);
  TF_AbstractOpSetOpType(fn_op, fn_name.c_str(), status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

                      absl::Span<AbstractTensorHandle* const> inputs,
                      absl::Span<AbstractTensorHandle*> outputs,
                      bool use_function) {
  AbstractTensorHandle* model_outputs[1];

  // Run the model.
  TF_RETURN_IF_ERROR(
      RunModel(forward, ctx, inputs, model_outputs, use_function));
  AbstractTensorHandlePtr model_out(model_outputs[0]);

  TF_Tensor* model_out_tensor;