TFE_ContextSetExecutorForThread(context, executor_.get());
    op_.reset(TFE_NewOp(context, operation_name, status));
    if (TF_GetCode(status) != TF_OK) return;
    TFE_OpSetDevice(op_.get(), device_.c_str(), status);
    if (TF_GetCode(status) != TF_OK) return;
  } else {
    TFE_OpReset(op_.get(), operation_name, device_.c_str(), status);
    if (TF_GetCode(status) != TF_OK) return;
  }

    std::vector<MaybeParallelTensorUnowned> inputs, const char* operation_name,
    const TFE_OpAttrs* attributes, int expected_max_outputs,
    TF_Status* status) {
  absl::optional<std::vector<MaybeParallelTensorOwned>> result;
  // TODO(allenl): We should remove "TPU" from these op names at the very least,
  // or consider other ways of packing/unpacking parallel tensors.
  if (operation_name == std::string("TPUReplicatedInput")) {

  // if sanity checks on dtypes/metadata fail.
  absl::optional<std::vector<std::unique_ptr<ParallelTensor>>> Execute(
      TFE_Context* context, const std::vector<ParallelTensor*>& inputs,
      const char* operation_name, const TFE_OpAttrs* attributes,
      int expected_max_outputs, TF_Status* status) const;

  // A non-blocking version of `Execute`. After each call, `Join` must be called