// 3. Replace worker1 with a new worker in server_def_0 and server_def_1.
  ReplaceTaskInServerDef(&server_def_0, /*task_index=*/1, "localhost", port);
  ReplaceTaskInServerDef(&server_def_1, /*task_index=*/1, "localhost", port);
  // 4. Start a new task to replace worker1.
  server_def_0.set_task_index(1);
  worker_server1.release();
  ASSERT_TRUE(tensorflow::GrpcServer::Create(

  // Add a client.
  job_def->mutable_tasks()->insert(
      {0, tensorflow::strings::StrCat(
              "localhost:", tensorflow::testing::PickUnusedPortOrDie())});

  tensorflow::JobDef* job_def2 = cluster_def->add_job();
  job_def2->set_name("worker");

  // Copy over `host:port` at `task_index`
  for (auto task : cluster_server_def.cluster().job(0).tasks()) {
    if (task.first == task_index) {

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

  const char* first_device = "/job:worker/replica:0/task:1/device:CPU:0";
  const char* second_device = "/job:worker/replica:0/task:2/device:CPU:0";
  const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
  std::array<const char*, 2> underlying_devices{first_device, second_device};

      "/job:worker/replica:0/task:1/device:CPU:0";
  const char new_local_device_name[] =
      "/job:worker/replica:0/task:0/device:CPU:0";

  auto* h0_task1_new = TFE_TensorHandleCopyToDevice(
      h0_task0_new, ctx, new_remote_device_name, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_DeleteTensorHandle(h0_task0_new);
  TFE_DeleteTensorHandle(h0_task1_new);

  // on all tasks in the cluster. This call internally coordinates with other
  // tasks to initialize the eager context and TF server for multi-client
  // execution.
  virtual Status EnableCollectiveOps(const ServerDef& server_def) = 0;

  // Check if the remote task is alive.
  virtual Status CheckRemoteAlive(const std::string& remote_task_name,
                                  bool* is_alive) = 0;

    return;
  }
  std::vector<tensorflow::CoordinatedTask> task_vec(tasks.length);
  auto* task_iter = static_cast<const tensorflow::CoordinatedTask*>(tasks.data);
  for (size_t i = 0; i < tasks.length; ++i) {
    task_vec[i].set_job_name(task_iter->job_name());
    task_vec[i].set_task_id(task_iter->task_id());
    ++task_iter;
  }
  auto results = coord_agent->GetTaskState(task_vec);
  if (!results.ok()) {

  const char task0_name[] = "/job:localhost/replica:0/task:0/device:CPU:0";
  const char task1_name[] = "/job:localhost/replica:0/task:1/device:CPU:0";
  const char task2_name[] = "/job:localhost/replica:0/task:2/device:CPU:0";

  // Create one variable per task.
  TFE_TensorHandle* h0 = TestVariable(ctx, 1.0, task1_name);
  TFE_TensorHandle* h1 = TestVariable(ctx, 2.0, task2_name);

    job_def->mutable_tasks()->insert(
        {i, tensorflow::strings::StrCat("localhost:", port)});
    LOG(INFO) << "Picked test port: " << port << " for job: " << job_name
              << ", task: " << i;
  }
  return server_def;
}

tensorflow::ServerDef GetServerDef(int num_tasks) {
  return GetServerDef("localhost", num_tasks);
}

tensorflow::ServerDef GetMultiClientServerDef(const std::string& job_name,

         * It has an added advantage since you can you easily switch to different hardware accelerators (cpu, gpu, tpu) as per the task.
         * All you need is a good internet connection and you are all set.
      * Try to build TF from sources by changing CPU optimization flags.
   
   *Please let us know if this helps.*
   
windows_comment: >

// custom devices will fail.
//
// `device_name` must not name an existing physical or custom device. It must
// follow the format:
//
//    /job:<name>/replica:<replica>/task:<task>/device:<type>:<device_num>
//
// If the device is successfully registered, `status` is set to TF_OK. Otherwise
// the device is not usable. In case of a bad status, `device.delete_device` is