Intentionally not using GPU

Ray Core
1

Hello, I’m new to ray but obviously I love this library!

I have a question. I want to ray.init() or @ray.remote, specifying only CPU resource with GPU machine. It’s because:

  • for my calculation the overhead to copy the numpy array to torch.Tensor array is not much gain in terms of speed.
  • I’m inside Docker and the ray’s GPU detection does not work correctly. (I’m not thinking to dig this to resolve for now - in a nutshell, torch.cuda.is_available() returns True outside ray but ray says it is False.)

So how I can do this? I’m specifying num_gpus=0 but ray.available_resources() still returns
{'object_store_memory': 5855249203.0, 'node:172.17.0.23': 1.0, 'accelerator_type:RTX': 1.0, 'memory': 11710498407.0, 'CPU': 10.0}
And I think I want to delete this accelerator_type:RTX because, obviously, I explicitly tell ray num_gpus = 0 !

Thank you.

2

cc @Alex

I believe in this case, accelerator_type:RTX': 1.0 shouldn’t do anything! But I agreed ideally, we should not detect accelerator type if num GPU is specified as 0.

3

Hmmm ok, thank you.
My error was saying that “deserializing from torch to numpy failed”, so I assumed that it casted numpy to torch somewhere inside even when I set num_gpus = 0. That’s why I guessed the accelerator did something. I can check that later.

4

If you think it is related to the accelerator type, can you create a Github issue?