Is there a way to specify a fractional gpu requirement for a task? In the example at Dask webpage, it specifies a single GPU.
from distributed import Client
client = Client('scheduler:8786')
data = [client.submit(load, fn) for fn in filenames]
processed = [client.submit(process, d, resources={'GPU': 1}) for d in data]
final = client.submit(aggregate, processed, resources={'MEMORY': 70e9})
It will be a waste to allocate one GPU for a small model.
Ray library has a way of doing this.
Yes, fractional resources are allowed.
In general, resources are arbitrary units, so if you instantiate the scheduler with "GPU=2" or some "foo=10", the scheduler will keep track of these resources during task assignment, but scheduler does not really know (or care) what "GPU" or "foo" is. When submitting tasks you can pass {"GPU": 0.25} and the scheduler will assign multiple tasks assigned to the GPU worker, For example, if the scheduler is told that each worker has "foo=10", then the scheduler will ask each worker to run concurrently up to ten tasks submitted with {"foo":1}.
