}

  tensorflow::shape_inference::ShapeHandle shape = ic->output(output.index);

  int rank = -1;
  if (ic->RankKnown(shape)) {
    rank = ic->Rank(shape);
  }

  if (num_dims != rank) {
    status->status = InvalidArgument("Expected rank is ", num_dims,
                                     " but actual rank is ", rank);
    return;
  }

  if (num_dims == 0) {
    // Output shape is a scalar.

  return wrap(unwrap(o)->outputs[i]);
}
void TF_OutputListPushBack(TF_OutputList* o, TF_AbstractTensor* tensor,
                           TF_Status* s) {
  unwrap(o)->outputs.push_back(unwrap(tensor));
}

void TF_AbstractOpSetOpType(TF_AbstractOp* op, const char* const op_type,
                            TF_Status* s) {
  tsl::Set_TF_Status_from_Status(
      s, unwrap(op)->Reset(op_type,

  const auto& shape = reader->GetVariableToShapeMap().at(name);
  int rank = shape.dims();
  if (num_dims != rank) {
    status->status = InvalidArgument("Expected rank is ", num_dims,
                                     " but actual rank is ", rank);
    return;
  }
  for (int i = 0; i < num_dims; i++) {
    dims[i] = shape.dim_size(i);
  }
}

  forward_op_->inputs.push_back(input);
  return absl::OkStatus();
}
Status AddInputList(AbstractOperation* op_,
                    absl::Span<AbstractTensorHandle* const> inputs,
                    ForwardOperation* forward_op_) {
  TF_RETURN_IF_ERROR(op_->AddInputList(inputs));
  for (auto input : inputs) {
    forward_op_->inputs.push_back(input);
  }
  return absl::OkStatus();
}

};

TEST_P(CppGradients, TestIdentityNGrad) {
  // This test is interesting because the current implementation of GradientTape
  // would return [0, 1] whereas we use build_default_zeros_grads=false here
  // so we get back [nullptr, 1].

  AbstractTensorHandlePtr x1;
  {
    AbstractTensorHandle* x1_raw = nullptr;
    status_ = TestScalarTensorHandle<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), 1.0f, &x1_raw);

    components.reserve(underlying_devices_.size());
    for (int j = 0; j < underlying_devices_.size(); ++j) {
      components.push_back(std::move(per_device_output_tensors[j][i]));
    }
    if (expected_output_shapes[i].IsFullyDefined()) {
      per_device_outputs.push_back(ParallelTensor::FromTensorHandles(
          *this, std::move(components),
          absl::Span<const int64_t>(expected_output_shapes[i].dim_sizes()),

    plugin_status.reserve(files.size());
    for (int i = 0; i < files.size(); i++)
      plugin_status.push_back(TF_NewStatus());
    result = ops_->paths_exist(filesystem_.get(), translated_names.data(),
                               files.size(), plugin_status.data());
    for (int i = 0; i < files.size(); i++) {
      status->push_back(StatusFromTF_Status(plugin_status[i]));
      TF_DeleteStatus(plugin_status[i]);
    }
  }

    for (int i = 0; i < noutputs; ++i) {
      output_names_vec.push_back(string(output_names[i]));
    }
  }

  // Process control output names.
  std::vector<string> control_output_names_vec;
  if (control_output_names) {
    control_output_names_vec.reserve(ncontrol_outputs);
    for (int i = 0; i < ncontrol_outputs; ++i) {
      control_output_names_vec.push_back(string(control_output_names[i]));
    }
  }

    TFE_OpAddInput(op, h, status);
    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);

    std::vector<TFE_TensorHandle*> result;
    result.push_back(nullptr);
    int num_retvals = 1;
    TFE_Execute(op, result.data(), &num_retvals, status);
    TFE_DeleteOp(op);
    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
    ASSERT_EQ(num_retvals, 1);

    if (TF_GetCode(status.get()) != TF_OK) {
      ADD_FAILURE() << tag << " -- " << TF_Message(status.get());
      continue;
    }
    TFE_DeleteTensorHandle(hdevice);
    // Copy back to CPU
    TFE_TensorHandle* hcopy =
        TFE_TensorHandleCopyToDevice(hdevice2, ctx, kCPUDevice, status.get());
    if (TF_GetCode(status.get()) != TF_OK) {
      ADD_FAILURE() << tag << " -- " << TF_Message(status.get());