(*config.mutable_device_count())["CPU"] = num_cpu_devices;

  // TODO(b/113217601): This is needed for EagerContext::runner_ to use a
  // threadpool, so that we avoid the possibility of running the runner_ in the
  // threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the

  // the status from `TFE_Execute` and returns outputs if the status is OK.
  std::vector<TensorHandlePtr> Join(TF_Status* status);

  // Block until all Ops finished running on the thread.
  void AsyncWait(TF_Status* status);

 private:
  void Run();

  void Execute(TFE_Context* context, const char* operation_name,
               std::vector<TFE_TensorHandle*> inputs,

  return nullptr;
}

// For TFE_CustomDevice::copy_tensor_from_device in the parallel device
// registration.
//
// Currently this is an error, and un-packing ParallelTensors must be performed
// explicitly by running a TPUReplicatedOutput operation on the parallel device.
//
// TODO(allenl): There are some use-cases that are only supported by copying to
// host at the moment (e.g. debug print on a tensor, .numpy(), etc.). We either

bool GraphErrorInjectionPass::enabled_ = false;

// 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

    TF_Status status;
    // Let this custom device choose the device to pin this op on if it
    // implements the pinning function.
    if (device_.shall_pin_to_this_device != nullptr) {
      return device_.shall_pin_to_this_device(tensorflow::wrap(op), &status);
    }
    return errors::Unimplemented("No custom device pinning implementation.");
  }

 private:
  TFE_Context* context_;
  TFE_CustomDevice device_;
  void* info_;