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_;

    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 {

             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();
      }
    }
  }