TF_Tensor* t = TFE_AllocateHostTensor(ctx, datatype, dims, num_dims, status);
  memcpy(TF_TensorData(t), data, TF_TensorByteSize(t));
  TFE_TensorHandle* th = TFE_NewTensorHandleFromTensor(ctx, t, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TF_DeleteTensor(t);
  TF_DeleteStatus(status);
  return th;
}

// Return a scalar tensor handle with given values.
template <class T, TF_DataType datatype>

bool GradientTape<Gradient, BackwardFunction, TapeTensor>::ShouldRecord(
    absl::Span<const int64_t> tensor_ids,
    absl::Span<const tensorflow::DataType> dtypes) const {
  CHECK_EQ(tensor_ids.size(), dtypes.size());
  for (int i = 0; i < tensor_ids.size(); ++i) {
    if (tensor_tape_.find(tensor_ids[i]) != tensor_tape_.end()) {
      if (IsDtypeTrainable(dtypes[i])) {
        return true;
      }

  // 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;

  // Get pointer to the coordination service agent instance.

  // context and TF server for multi-client execution.
  virtual Status EnableCollectiveOps(const ServerDef& server_def) = 0;

  // Set a distributed manager that helps set up, update, and check liveness
  // of member tasks in the cluster.
  virtual void SetDistributedManager(
      std::unique_ptr<ImmediateExecutionDistributedManager> distributed) = 0;