How to define `num_gpus` in `ray.remote()` while not explicitly adding `@ray.remote` above the target class

Ray Core
1

How severe does this issue affect your experience of using Ray?

  • Medium: It contributes to significant difficulty to complete my task, but I can work around it.

My question is: How to define num_gpus in ray.remote() while not explicitly adding @ray.remote above the target class?

I’m trying to use ray to accelerate my federated learning training task. Here are some codes about the implementation:

class FLbenchTrainer:
    def __init__(
        self, server, client_cls, mode: str, num_workers: int, init_args: Dict
    ):
        self.server = server
        self.mode = mode
        self.num_workers = num_workers

        if self.mode == "serial":
            self.worker = client_cls(**init_args)
        elif self.mode == "parallel":
            ray_client = ray.remote(client_cls)
            self.workers: List[ActorHandle] = [
                ray_client.remote(**init_args) for _ in range(self.num_workers)
            ]

For some reasons, I don’t want to add @ray.remote decorator above all the client_cls in my project explicitly. Because sometimes I need my code run in serial. So I callray.remote(client_cls) to get the Ray Actor.

Since the num_gpus is not set, when my num_workers > 1 (equals to the num of ray actors), no actor can utilize the gpu resource properly.

I know I need to set num_gpus = NUM_MY_GPU / num_workers to allocate actors their own gpu resource. But like I said, the num_workers is not fixed and sometimes I don’t want ray activated, I don’t want to add @ray.remote explicitly. So what should I do?

1 Like
2

You can pass num_gpus to the options() wrapper, like:

class MyActor:
  pass

remote_cls = ray.remote(MyActor)

remote_class_1_gpu = remote_cls.options(num_gpus=1)
remote_class_4_gpu = remote_cls.options(num_gpus=4)
4

Thx for your solution