Maybe Inconsistent GPU Utilization Due to Hardcoded `num_gpus_per_actor` and `num_gpus`

https://github.com/THUDM/slime/blob/921e2656aa3e5f7ff9fefe7cccda962a8bacfc13/slime/ray/rollout.py#L215-L226
https://github.com/THUDM/slime/blob/921e2656aa3e5f7ff9fefe7cccda962a8bacfc13/slime/ray/placement_group.py#L112-L120

Description:
The current implementation hardcodes num_gpus_per_actor=0.4 and num_gpus=0.2 in the GPU resource allocation logic. This may lead to significant underutilization of GPU resources in multi-node/multi-GPU setups？

Example Scenario:

• Cluster: 4 nodes × 8 GPUs = 32 GPUs total.
• Actual utilization:
  • Rollout actors: 0.4 GPU/actor × 32 GPUs → 12.8 GPUs
  • SGLang engine: 0.2 GPU/engine × 4 engines → 0.8 GPUs
  • Total used: 13.6 GPUs (42.5% utilization)

Code Snippets:

1. Hardcoded values in RolloutRayActor and allocate_train_group:

# Rollout engine setup
rollout_engines.append(
    RolloutRayActor.options(
        num_cpus=num_cpus,
        num_gpus=num_gpus,  # Hardcoded to 0.2
        scheduling_strategy=PlacementGroupSchedulingStrategy(...)
    )
)

# Train group allocation
def allocate_train_group(...):
    return RayTrainGroup(
        ...,
        num_gpus_per_actor=0.4,  # Hardcoded value
    )

Question

Does this configuration lead to inefficient GPU utilization? Should we make num_gpus_per_actor=0.4 and num_gpus=0.2 configurable parameters instead of hardcoding them in the code?

Read slime/ray/rollout.py at lines 215-226 and slime/ray/placement_group.py at lines 112-120, then trace how the hardcoded GPU values are used for rollout actors and train-group allocation. Confirm the intended allocation across the described multi-node setup; done should make the values configurable and demonstrate that the resulting resource usage matches the configuration.