No trial resources are available for launching the actor

RLlib
1

High

  • None: Just asking a question out of curiosity
  • Low: It annoys or frustrates me for a moment.
  • Medium: It contributes to significant difficulty to complete my task, but I can work around it.
  • High: It blocks me to complete my task.

My code:

trainable_with_resources = tune.with_resources(
    experiment_fn,
    resources={
        "gpu": 1, 
        "cpu": 4}
    )

tuner = tune.Tuner(
    trainable_with_resources,
    tune_config=tune.TuneConfig(
        num_samples=20,
        scheduler=ASHAScheduler(metric="eval_reward", mode="max"),
    ),
    param_space=search_space
)
results = tuner.fit()

Error:

No trial resources are available for launching the {submitted} `{name}`. "
ray.tune.error.TuneError: No trial resources are available for launching the actor `ray.rllib.evaluation.rollout_worker.RolloutWorker.__init__`. To resolve this, specify the Tune option:

>  resources_per_trial=tune.PlacementGroupFactory(
>    [{'GPU': 1.0, 'CPU': 4.0}] + [{'CPU': 1.0}] * N
>  )

In error traceback it is said that i could find solution here: A Guide To Parallelism and Resources — Ray 2.1.0

But i can not find one.

2

Does your trainable also launch other remote actors? It seems so. From error message, there are also rollout workers.
The current way you have is asking Ray to allocate gpu:1 + cpu: 4 to the trainable head and thus there is none left for rollout workers.
How to fix it:
Maybe you could take a look at rllib/examples/custom_experiment.py, especially this line

tune.with_resources(experiment, ppo.PPO.default_resource_request(config)),
3

In other words i allocate 1 gpu for trainable actor only. Rest 4 CPUs will be used by rollout workers.
Trainable fits model
Rollout workers make samples for it. Right?

I tried to allocate 1 gpu only same error.

Im not sure where i can find this file, because i run everything in colab botebook.

4

here is the line: ray/custom_experiment.py at master · ray-project/ray · GitHub

1 Like
5

It works poerfect. But it tells me that
2022-11-14 02:46:16,788 WARNING insufficient_resources_manager.py:128 -- Ignore this message if the cluster is autoscaling. You asked for 1.0 cpu and 1.9999999999999998 gpu per trial

Sounds like i still have three questions:

  1. How can i fix number of requested resources or read smth about resources_per_trial or num_workers for rllib.
  2. How can i specify number of trials for hyperparameter search. Right now only experiment seems to be on
  3. How can i add ASHA (hyperparameter improvement)
    Getting Started — Ray 2.1.0

p.s. where can i find smth aboout tune.with resources (i didnt find anything in docs)