* could actually hurt in some cases, as it forces us to keep all inputs in memory until the
       * final input completes.
       */
      @RetainedLocalRef
      ImmutableCollection<? extends ListenableFuture<? extends InputT>> localFutures = futures;
      ImmutableCollection<? extends Future<? extends InputT>> localFuturesOrNull =
          collectsValues ? localFutures : null;

    for (int i = 0; i < inputs.size(); i++) {
      GraphTensor* t = dyn_cast<GraphTensor>(inputs[i]);
      if (!t) {
        return tensorflow::errors::InvalidArgument(
            "Unable to cast input to GraphTensor");
      }
      tf_outputs[i] = t->output_;
    }
    TF_AddInputList(op_.get(), tf_outputs.data(), tf_outputs.size());
    return absl::OkStatus();
  }

    TF_EXCLUSIVE_LOCKS_REQUIRED(fn_body->mu) {
  input_tensors->reserve(ninputs);
  for (int i = 0; i < ninputs; ++i) {
    Node* node = inputs[i].oper ? &inputs[i].oper->node : nullptr;
    int idx = inputs[i].index;

    TF_RETURN_WITH_CONTEXT_IF_ERROR(
        fn_body->graph.IsValidOutputTensor(node, idx),
        "Encountered while processing input ", i, " into function '", fn_name,
        "'");

void TF_AddInput(TF_OperationDescription* desc, TF_Output input) {
  desc->node_builder.Input(&input.oper->node, input.index);
}

void TF_AddInputList(TF_OperationDescription* desc, const TF_Output* inputs,
                     int num_inputs) {
  std::vector<NodeBuilder::NodeOut> input_list;
  input_list.reserve(num_inputs);
  for (int i = 0; i < num_inputs; ++i) {
    input_list.emplace_back(&inputs[i].oper->node, inputs[i].index);
  }

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

  // Build inputs and outputs.
  TF_AbstractTensor* inputs[2] = {at, at};
  TF_OutputList* o = TF_NewOutputList();
  TF_OutputListSetNumOutputs(o, 1, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  // Execute.
  TF_ExecuteOperation(op, 2, inputs, o, status.get());

  parallel_inputs.reserve(inputs.size());
  implicitly_broadcast_tensors.reserve(inputs.size());  // not tight
  for (const auto& input : inputs) {
    if (absl::holds_alternative<TFE_TensorHandle*>(input)) {
      if (operation_name == std::string("_EagerConst")) {
        // Non-parallel tensors from _EagerConst/tf.constant are implicitly
        // broadcast, i.e. set as the input to each parallel operation. This

        "Unable to find the placeholder node as input to enqueue in the TF "
        "graph.");
    return;
  }

  VLOG(1) << "Running the enqueue op";
  TF_Output input{placeholder_op, 0};
  TF_SessionRun(session, /*run_options*/ nullptr,
                // input related parameters
                /*inputs*/ &input, /*input_values*/ &tensor, /*ninputs*/ 1,
                // output related parameters

      TF_AddGradientsWithPrefix(graph_, prefix, outputs, noutputs, inputs,
                                ninputs, grad_inputs, s_, grad_outputs);
    } else {
      TF_AddGradientsWithPrefix(graph_, prefix, outputs, noutputs, inputs,
                                ninputs, nullptr, s_, grad_outputs);
    }
  }

  void BuildErrorGraph(TF_Output* inputs, TF_Output* outputs) {

      for (ClosingFuture<?> input : inputs) {
        input.becomeSubsumedInto(closeables);
      }
    }

    /**
     * Returns a new {@code ClosingFuture} pipeline step derived from the inputs by applying a
     * combining function to their values. The function can use a {@link DeferredCloser} to capture
     * objects to be closed when the pipeline is done.
     *

  for (auto input : inputs) {
    forward_op_->inputs.push_back(input);
  }
  return absl::OkStatus();
}

absl::Status SetAttrString(AbstractOperation* op_, const char* attr_name,
                           const char* data, size_t length,
                           ForwardOperation* forward_op_) {
  forward_op_->attrs.Set(attr_name, StringPiece(data, length));