[train_async.py] ValueError: Invalid type of object refs, <class 'NoneType'>, is given. 'object_refs' must either be an ObjectRef or a list of ObjectRefs.

您好，很荣幸看到slime这么棒的工作。
我对slime框架的使用很感兴趣，目前正在尝试slime中关于自定义奖励函数的用法，实验环境为4*H100，模型为Qwen3-4B进行GRPO训练，其中设置
-- python3 train_async.py
--actor-num-nodes 1
--actor-num-gpus-per-node 2
--rollout-num-gpus 1
采用train_async.py 训推分离的模式，另外一张gpu用来本地部署模型作为奖励函数的api调用。

我已经做好前期的准备工作，并且调整合适的rollout_batch_size等参数可以运行起来，
但是当num_rollout训练完成保存模型后，日志打印报错信息如下：
(MegatronTrainRayActor pid=2121369) import pynvml # type: ignore[import]
Traceback (most recent call last):
File "/slime/train_async.py", line 108, in
train(args)
File "/slime/train_async.py", line 95, in train
rollout_data_curr_ref = ray.get(rollout_data_next_future)
File "/root/micromamba/envs/slime/lib/python3.12/site-packages/ray/_private/auto_init_hook.py, line 22, in auto_init_wrapper
return fn(*args, **kwargs)
File "/root/micromamba/envs/slime/lib/python3.12/site-packages/ray/client_mode_hook.py, line 104, in wrapper
return func(*args, **kwargs)
File "/root/micromamba/envs/slime/lib/python3.12/site-packages/ray/_private/worker.py, line 2877, in get
raise ValueError(
ValueError: Invalid type of object refs, <class 'NoneType'>, is given. 'object_refs' must either be an ObjectRef or a list of ObjectRefs.

Status message: Job entrypoint command failed with exit code 1, last available logs (truncated to 20,000 chars):
File "/root/micromamba/envs/slime/lib/python3.12/site-packages/ray/client_mode_hook.py, line 104, in wrapper
return func(*args, **kwargs)
File "/root/micromamba/envs/slime/lib/python3.12/site-packages/ray/_private/worker.py, line 2877, in get
raise ValueError(
ValueError: Invalid type of object refs, <class 'NoneType'>, is given. 'object_refs' must either be an ObjectRef or a list of ObjectRefs.

这个问题困扰了我很长时间，请教各位大佬有没有遇到过类似的问题？或者这个问题应该如何解决？
十分感谢

Start in train_async.py at lines 95 and 108, then trace how rollout_data_next_future is produced before it is passed to ray.get. Reproduce the reported failure with the stated train_async.py configuration and inspect why the value is None; done means the training loop receives a valid object reference or reports the underlying failure clearly.