.ok());
  ASSERT_TRUE(worker_server1->Start().ok());

  server_def.set_task_index(2);
  std::unique_ptr<tensorflow::GrpcServer> worker_server2;
  ASSERT_TRUE(tensorflow::GrpcServer::Create(
                  server_def, tensorflow::Env::Default(), &worker_server2)
                  .ok());
  ASSERT_TRUE(worker_server2->Start().ok());

  TFE_ContextSetServerDef(context.get(), 0, serialized.data(),

#if defined(IS_MOBILE_PLATFORM) || defined(IS_SLIM_BUILD)
  status->status = tensorflow::errors::Unimplemented(
      "Server functionality is not supported on mobile");
#else
  status->status = server->server->Start();
#endif  // defined(IS_MOBILE_PLATFORM) || defined(IS_SLIM_BUILD)
}

void TF_ServerStop(TF_Server* server, TF_Status* status) {
#if defined(IS_MOBILE_PLATFORM) || defined(IS_SLIM_BUILD)

                                    /*remote_func_outputs=*/true);
}
TEST(CAPI, RemoteExecuteSilentCopiesLocalAsyncFuncOrdering) {
  // A remote input may be not ready when we start running a function. Test that
  // the function execution should wait until the remote input is ready.
  TestRemoteExecuteSilentCopiesFunc(/*async=*/true, /*remote=*/false,

namespace gradients {

using namespace std;

// ================== Helper functions =================

// Fills data with values [start,end) with given step size.
void Range(vector<int32_t>* data, int32_t start, int32_t end,
           int32_t step = 1) {
  for (int32_t i = start; i < end; i += step) {
    (*data)[i] = i;
  }
}

// Fills out_dims with the dimensions of the given tensor.

                  .ok());
  ASSERT_TRUE(worker_server1->Start().ok());

  server_def.set_task_index(2);
  std::unique_ptr<tensorflow::GrpcServer> worker_server2;
  ASSERT_TRUE(tensorflow::GrpcServer::Create(
                  server_def, tensorflow::Env::Default(), &worker_server2)
                  .ok());
  ASSERT_TRUE(worker_server2->Start().ok());

  TF_Status* status = TF_NewStatus();

  tensorflow::mutex execution_mutex_;
  ExecutionState execution_state_ TF_GUARDED_BY(execution_mutex_) =
      ExecutionState::kIdle;
  // Tells the worker thread that there is new work.
  tensorflow::condition_variable start_execute_;
  // The worker thread notifies that work has finished.
  tensorflow::condition_variable finished_execute_;
  // Notifies a StartExecute that the previous Join has finished.

                         status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  // Create a variable handle (uninitialized to start) placed on the parallel
  // device.
  std::function<void(Variable*)> variable_deleter = [&](Variable* to_delete) {
    to_delete->Destroy(context, status.get());

  ASSERT_TRUE(tensorflow::GrpcServer::Create(
                  server_def, tensorflow::Env::Default(), &worker_server)
                  .ok());
  ASSERT_TRUE(worker_server->Start().ok());

  TF_Status* status = TF_NewStatus();
  TFE_ContextOptions* opts = TFE_NewContextOptions();
  TFE_ContextOptionsSetAsync(opts, static_cast<unsigned char>(async));

  ASSERT_TRUE(tensorflow::GrpcServer::Create(
                  server_def, tensorflow::Env::Default(), &worker_server)
                  .ok());
  ASSERT_TRUE(worker_server->Start().ok());

  TF_Status* status = TF_NewStatus();
  TFE_ContextOptions* opts = TFE_NewContextOptions();
  TFE_ContextOptionsSetAsync(opts, static_cast<unsigned char>(async));
  TFE_ContextOptionsSetDevicePlacementPolicy(opts,

  }
  // Calculate the block-aligned start and end of the read.
  size_t start = block_size_ * (offset / block_size_);
  size_t finish = block_size_ * ((offset + n) / block_size_);
  if (finish < offset + n) {
    finish += block_size_;
  }
  size_t total_bytes_transferred = 0;
  // Now iterate through the blocks, reading them one at a time.
  for (size_t pos = start; pos < finish; pos += block_size_) {