TFE_DeleteContextOptions(opts);
    TFE_Executor* executor = TFE_NewExecutor(
        /*is_async=*/false, /*enable_streaming_enqueue=*/true,
        /*in_flight_nodes_limit=*/0);
    TFE_ContextSetExecutorForThread(ctx, executor);

    TFE_DeleteContext(ctx);
    TFE_DeleteExecutor(executor);
  }

  {
    TFE_ContextOptions* opts = TFE_NewContextOptions();

                                                        TF_Buffer* buf,
                                                        TF_Status* status);

// Some TF ops need a step container to be set to limit the lifetime of some
// resources (mostly TensorArray and Stack, used in while loop gradients in
// graph mode). Calling this on a context tells it to start a step.
TF_CAPI_EXPORT extern void TFE_ContextStartStep(TFE_Context* ctx);

// executed in sequence. Assigning nodes to different executors allows executing
// nodes in parallel.
// in_flight_nodes_limit: when is_async is true, this value controls the
// maximum number of in flight async nodes. Enqueuing of additional async ops
// after the limit is reached blocks until some inflight nodes finishes.
// The effect is bounding the memory held by inflight TensorHandles that are

    return errors::InvalidArgument("Value specified for `", attr_name, "` has ",
                                   num_dims,
                                   " dimensions which is over the limit of ",
                                   TensorShape::MaxDimensions(), ".");
  }
  TensorShapeProto proto;
  if (num_dims < 0) {
    proto.set_unknown_rank(true);
  } else {

  status->status = absl::OkStatus();
}

TFE_Executor* TFE_NewExecutor(bool is_async, bool enable_streaming_enqueue,
                              int in_flight_nodes_limit) {
  return new TFE_Executor(is_async, enable_streaming_enqueue,
                          in_flight_nodes_limit);
}

void TFE_DeleteExecutor(TFE_Executor* executor) { delete executor; }

bool TFE_ExecutorIsAsync(TFE_Executor* executor) {

    auto virtual_devices =
        gpu_options->mutable_experimental()->add_virtual_devices();
    for (int i = 0; i < num_virtual_gpus; ++i) {
      virtual_devices->add_memory_limit_mb(200);
    }
  }
  return server_def;
}

TFE_TensorHandle* CreateVarHandle(TFE_Context* ctx,
                                  const tensorflow::string& device_name,

        // parallel device.
        device_(device),
        executor_(
            TFE_NewExecutor(is_async, /*enable_streaming_enqueue=*/true,
                            /*in_flight_nodes_limit=*/in_flight_nodes_limit)),
        op_(nullptr),
        thread_(tensorflow::Env::Default()->StartThread(
            tensorflow::ThreadOptions(), "parallel_device_execute",
            std::bind(&DeviceThread::Run, this))) {}

  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ~ParallelDevice();

  // Helper to copy a tensor handle from another device once for each component
  // of the ParallelDevice.
  //
  // Sets a bad status and returns a nullptr if `tensor` is already on the