input_tensors->emplace_back(node, idx);

    const auto& iter = input_nodes->find(node);
    if (iter == input_nodes->end()) {
      input_nodes->insert({node, {idx}});
    } else {
      auto& indices = iter->second;
      if (std::find(indices.begin(), indices.end(), idx) != indices.end()) {
        return InvalidArgument("TF_Output ", node->name(), ":", idx,
                               " appears more than once in the input list");

                    /*input_shapes*/ {make_shape({3, 2}), make_shape({2, 4})},
                    /*input_tensors*/ {},
                    /*expected_shape*/ make_shape({3, 4}));

  // Infer shape when second operand has unknown shape.
  CheckOutputShapes(matmul_op,
                    /*input_shapes*/ {make_shape({3, 2}), unknown_shape()},
                    /*input_tensors*/ {},

    TF_DeleteAbstractOp(add_op);
    // Extract the resulting tensor.
    add_output1 = TF_OutputListGet(add_outputs, 0);
    TF_DeleteOutputList(add_outputs);
  }

  // Same with a second "Add" computing `arg1 + arg1`.
  TF_AbstractTensor* add_output2;
  {
    // Build an abstract operation, inputs and output.
    auto* add_op = TF_NewAbstractOp(graph_ctx);
    TF_AbstractOpSetOpType(add_op, "Add", s);

  RegisterParallelDevice(context.get(), first_device_name,
                         first_underlying_devices, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Create a second parallel device with the first parallel device and one
  // additional CPU.
  const char* second_device_name =
      "/job:localhost/replica:0/task:0/device:CUSTOM:1";
  std::array<const char*, 2> second_underlying_devices{

  if (iter == registry_.end()) {
    const string error_msg = "No gradient defined for op: " + op.op_name + ".";
    return errors::NotFound(error_msg);
  }
  gradient_function->reset(iter->second(op));
  return absl::OkStatus();
}

TapeTensor::TapeTensor(AbstractTensorHandle* handle) : handle_(handle) {
  handle_->Ref();
}
TapeTensor::TapeTensor(const TapeTensor& other) {
  handle_ = other.handle_;

  TFE_Op* func_op = TFE_NewOp(ctx, func.name().data(), status);
  for (const auto& attr : func.attr()) {
    if (!status->status.ok()) return nullptr;
    SetOpAttrValueScalar(ctx, func_op, attr.second, attr.first.data(), status);
    if (!status->status.ok()) return nullptr;
  }
  return func_op;
}
}  // namespace

void TFE_ContextStartStep(TFE_Context* ctx) {

    if (attr.first == "dtype") {
      if (attr.second.type() == tensorflow::DT_INT32) {
        found_dtype = true;
      } else {
        return false;
      }
    } else if (attr.first == "value") {
      if (attr.second.has_tensor() &&
          attr.second.tensor().int_val_size() == 1 &&
          attr.second.tensor().int_val(0) == v) {
        found_value = true;
      } else {

// Test to ensure that a registered graph optimization pass is only executed
// once (i.e., on the main function side) in running distributed functions.
// This test creates a cluster with two workers, create a variable on the
// second worker, and run a distributed function (VariableAddFunction) whose ops
// span the local and remote workers. If the graph optimization pass is executed
// on both the main function side and the component function side, an error will