ops_->flush(file_.get(), plugin_status.get());
  return StatusFromTF_Status(plugin_status.get());
}

Status ModularWritableFile::Sync() {
  if (ops_->sync == nullptr) return Flush();

  UniquePtrTo_TF_Status plugin_status(TF_NewStatus(), TF_DeleteStatus);
  ops_->sync(file_.get(), plugin_status.get());
  return StatusFromTF_Status(plugin_status.get());
}

  if (forward_input) {
    TFE_DeleteTensorHandle(n);
  }

  int num_retvals = 1;
  if (async) {
    // Enqueue dummy ops so we backlog async execution & actually test async.
    // This is usually unnecessary, but we've experienced the occasional test
    // failure when testing async mode with no explicit forwarding.
    for (int i = 0; i < 100000; ++i) {
      TFE_Op* add_op_dummy = AddOp(ctx, m, m);

  TFE_DEVICE_PLACEMENT_SILENT_FOR_INT32 = 3,
} TFE_ContextDevicePlacementPolicy;
// LINT.ThenChange(//tensorflow/c/eager/immediate_execution_context.h)

// Sets the default execution mode (sync/async). Note that this can be
// overridden per thread using TFE_ContextSetExecutorForThread.
TF_CAPI_EXPORT extern void TFE_ContextOptionsSetAsync(TFE_ContextOptions*,

  ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
  TF_DeleteFunction(fn);

  for (bool async : {false, true, false}) {
    TFE_Executor* old_executor = TFE_ContextGetExecutorForThread(ctx);
    TFE_Executor* executor = TFE_NewExecutor(
        /*is_async=*/async, /*enable_streaming_enqueue=*/true,
        /*in_flight_nodes_limit=*/0);
    TFE_ContextSetExecutorForThread(ctx, executor);

// which creates a remote input, to simulate a scenario that the remote input
// 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);

            gcs_file->object.c_str());
    if (TF_GetCode(status) != TF_OK) return;
    gcs_file->sync_need = false;
  } else {
    TF_SetStatus(status, TF_OK, "");
  }
}

void Sync(const TF_WritableFile* file, TF_Status* status) {
  auto gcs_file = static_cast<GCSFile*>(file->plugin_file);
  TF_VLog(3, "Sync: gs://%s/%s", gcs_file->bucket.c_str(),
          gcs_file->object.c_str());
  Flush(file, status);
}

  auto* cc_file = reinterpret_cast<::tensorflow::WritableFile*>(handle);
  TF_SetStatus(status, TF_OK, "");
  ::tensorflow::Set_TF_Status_from_Status(status, cc_file->Sync());
}

void TF_FlushWritableFile(TF_WritableFileHandle* handle, TF_Status* status) {
  auto* cc_file = reinterpret_cast<::tensorflow::WritableFile*>(handle);
  TF_SetStatus(status, TF_OK, "");

  if (fflush(posix_file->handle) != 0)
    TF_SetStatusFromIOError(status, errno, posix_file->filename);
}

static void Sync(const TF_WritableFile* file, TF_Status* status) {
  // For historical reasons, this does the same as `Flush` at the moment.
  // TODO(b/144055243): This should use `fsync`/`sync`.
  Flush(file, status);
}

static void Close(const TF_WritableFile* file, TF_Status* status) {

  ~ModularWritableFile() override { ops_->cleanup(file_.get()); }

  Status Append(StringPiece data) override;
  Status Close() override;
  Status Flush() override;
  Status Sync() override;
  Status Name(StringPiece* result) const override;
  Status Tell(int64_t* position) override;

 private:
  std::string filename_;
  std::unique_ptr<TF_WritableFile> file_;

// executed in sequence. Assigning nodes to different executors allows executing
// nodes in parallel.
// in_flight_nodes_limit: when is_async is true, this value controls the
// maximum number of in flight async nodes. Enqueuing of additional async ops
// after the limit is reached blocks until some inflight nodes finishes.
// The effect is bounding the memory held by inflight TensorHandles that are