ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  std::vector<TensorHandlePtr> vector_handles;
  vector_handles.reserve(2);
  vector_handles.push_back(std::move(two_vector));
  vector_handles.push_back(std::move(three_vector));
  std::unique_ptr<ParallelTensor> unknown_length_vector =
      ParallelTensor::FromTensorHandles(
          parallel_device, std::move(vector_handles), status.get());

    all_input_tensors.reserve(num_inputs);
    for (int i = 0; i < num_inputs; ++i) {
      if (input_tensors[i] == nullptr) continue;
      all_input_tensors.emplace_back();
      Tensor& input_tensor = all_input_tensors.back();
      status->status = TF_TensorToTensor(input_tensors[i], &input_tensor);
      if (!status->status.ok()) return;
      input_tensors_vector[i] = &input_tensor;
    }
  }

  inputs_.clear();
  for (const auto& p : inputs) {
    inputs_.emplace_back(TF_Output{p.first, 0});
    input_values_.emplace_back(p.second);
  }
}

void CSession::SetOutputs(std::initializer_list<TF_Operation*> outputs) {
  ResetOutputValues();
  outputs_.clear();
  for (TF_Operation* o : outputs) {
    outputs_.emplace_back(TF_Output{o, 0});
  }
  output_values_.resize(outputs_.size());
}

  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();
}

        return result;
      }
      components.emplace_back(TFE_TensorHandleCopySharingTensor(
          absl::get<TFE_TensorHandle*>(inputs[i]), status));
    }
    std::vector<MaybeParallelTensorOwned> result_content;
    result_content.reserve(1);
    result_content.push_back(ParallelTensor::FromTensorHandles(
        parallel_device, std::move(components), status));

    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()),

    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]));
    }
  }

      Multiply(context.get(), second_combined_value.get(),
               second_negative_one.get(), status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Un-pack the parallel tensor to verify that the operation was
  // successful. The resulting structure should be:
  //   second_device{first_device{1. * 3., 2. * 3.}, 3. * 3.}.
  std::array<TensorHandlePtr, 2> second_components;

                   &FloatDeallocator, nullptr),
      TF_DeleteTensor);
  return TensorHandlePtr(TFE_NewTensorHandle(tensor.get(), status));
}

// Helper to un-pack `num_replicas` TFE_TensorHandles from one parallel handle.
template <std::size_t num_replicas>
void ExtractPerDeviceValues(
    TFE_Context* context, TFE_TensorHandle* input,

                          TF_Status* status) {
  std::vector<tensorflow::VariableInfo*> variable_ptrs;
  variable_ptrs.reserve(num_vars);
  for (int i = 0; i < num_vars; ++i) {
    variable_ptrs.push_back(&(vars[i]->var_info));
  }
  absl::Status cc_status = LockVariables(absl::MakeSpan(variable_ptrs));
  status->status = cc_status;
}

void TF_AllocateTempForVariableInfo(TF_OpKernelContext* ctx,