[<Ray component: util>] PlacementGroup bundle index not effect when num_gpus<1

[Hanqer]

What happened + What you expected to happen

import ray 
import os 

from ray.util.placement_group import PlacementGroup, PlacementGroupSchedulingStrategy, placement_group

ray.init() 

@ray.remote(num_gpus=1)
class Test():
    def __init__(self, id):
        self.id = id 

    def test(self):
        print(f"Actor ID: {self.id}, CUDA_VISIBLE_DEVICES: {os.environ['CUDA_VISIBLE_DEVICES']}")

if __name__ == "__main__":
    pg = placement_group([{"GPU": 1, "CPU": 1} for _ in range(4)])
    ray.get(pg.ready())

    actors = [
        Test.options(
            num_gpus=0.5,
            num_cpus=0.5,
            scheduling_strategy=PlacementGroupSchedulingStrategy(
                placement_group=pg, placement_group_bundle_index=i
            ),
        ).remote(i)
        for i in range(4)]

    for actor in actors:
        ray.get(actor.test.remote())

I think this should output the cuda_devices: 0 1 2 3, since each actor requires 0.5 gpus, and each bundle has 1 gpu resouces. Total four actors are evenly require bundle with placement_group_bundle_index of bundle list with four items. I holp this code place each actor on 0.5 gpus but with different 4 devieces. However, this result in two actors on gpu devices 0, and two actors on gpu devices 1. I want to knwo why, thx!

Versions / Dependencies

ray==2.12.0

Reproduction script

import ray 
import os 

from ray.util.placement_group import PlacementGroup, PlacementGroupSchedulingStrategy, placement_group

ray.init() 

@ray.remote(num_gpus=1)
class Test():
    def __init__(self, id):
        self.id = id 

    def test(self):
        print(f"Actor ID: {self.id}, CUDA_VISIBLE_DEVICES: {os.environ['CUDA_VISIBLE_DEVICES']}")

if __name__ == "__main__":
    pg = placement_group([{"GPU": 1, "CPU": 1} for _ in range(4)])
    ray.get(pg.ready())

    actors = [
        Test.options(
            num_gpus=0.5,
            num_cpus=0.5,
            scheduling_strategy=PlacementGroupSchedulingStrategy(
                placement_group=pg, placement_group_bundle_index=i
            ),
        ).remote(i)
        for i in range(4)]

    for actor in actors:
        ray.get(actor.test.remote())

Issue Severity

High: It blocks me from completing my task.

[rynewang]

This has been fixed in master in #48088 (not released yet, aim ray 2.40)

(Test pid=99523) Actor ID: 0, CUDA_VISIBLE_DEVICES: 0
(Test pid=99520) Actor ID: 1, CUDA_VISIBLE_DEVICES: 1
(Test pid=99522) Actor ID: 2, CUDA_VISIBLE_DEVICES: 2
(Test pid=99525) Actor ID: 3, CUDA_VISIBLE_DEVICES: 3