device_threads_.emplace_back(new DeviceThread(
        devices[device_index].c_str(), is_async, in_flight_nodes_limit));
  }
}

// Necessary for a unique_ptr to a forward-declared type.
ParallelDevice::~ParallelDevice() = default;

std::unique_ptr<ParallelTensor> ParallelDevice::CopyToParallelDevice(
    TFE_Context* context, TFE_TensorHandle* tensor, TF_Status* status) const {

    forward_grads.resize(output_tensors.size());
    TF_RETURN_IF_ERROR(ForwardpropFromTape(
        op_type, output_tensors, backward_function_getter,
        backward_function_deleter, in_grads, absl::MakeSpan(forward_grads)));
  } else {
    TF_RETURN_IF_ERROR(
        (*forward_function)(in_grads, &forward_grads, use_batch_));
  }
  for (int i = 0; i < forward_grads.size(); ++i) {
    if (forward_grads[i] != nullptr) {

      ASSERT_EQ(errors::OK, status_.code()) << status_.message();
      immediate_execution_ctx_.reset(ctx_raw);
    }

    // Computing numerical gradients with TensorFloat-32 is numerically
    // unstable. Some forward pass tests also fail with TensorFloat-32 due to
    // low tolerances
    enable_tensor_float_32_execution(false);
  }

  AbstractContextPtr immediate_execution_ctx_;
  GradientRegistry registry_;

             const char* raw_device_name, ForwardOperation* forward_op_) {
  forward_op_->op_name = op;
  forward_op_->attrs.Reset(op);
  return op_->Reset(op, raw_device_name);
}
Status AddInput(AbstractOperation* op_, AbstractTensorHandle* input,
                ForwardOperation* forward_op_) {
  TF_RETURN_IF_ERROR(op_->AddInput(input));
  forward_op_->inputs.push_back(input);
  return absl::OkStatus();
}

                                    vector<AbstractTensorHandle*> f_outputs)
      : forward_inputs_(f_inputs), forward_outputs_(f_outputs) {
    for (auto input : forward_inputs_) {
      if (input) {
        input->Ref();
      }
    }
    for (auto output : forward_outputs_) {
      if (output) {
        output->Ref();
      }
    }
  }

    your response for more than 2 weeks.

**4. Approved**

-   Once the PR is approved, it gets `kokoro:force-run` label applied and it
    initiates CI/CD tests.
-   We can't move forward if these tests fail.
-   In such situations, we may request you to make further changes to your PR
    for the tests to pass.
-   Once the tests pass, we now bring all the code into the internal code base,