CUDA serialization error with Population Based tuning


Thank you for your great work!
I am encountering an issue with cuda serialization while using PBT.
I setup my tune.Tuner to have multiple workers (e.g., 4), and 1 gpu.
When .fit, I am able to generate checkpoints for the different workers, but at some point early on after (it seems) all trials have been explored for the first number of time_attr, I get the following error:

(RolloutWorker pid=73953) Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location=torch.device(‘cpu’) to map your storages to the CPU.
(RolloutWorker pid=73953) Traceback (most recent call last):
(RolloutWorker pid=73953) File “/usr/local/lib/python3.10/site-packages/ray/_private/”, line 404, in deserialize_objects
(RolloutWorker pid=73953) obj = self._deserialize_object(data, metadata, object_ref)
(RolloutWorker pid=73953) File “/usr/local/lib/python3.10/site-packages/ray/_private/”, line 270, in _deserialize_object
(RolloutWorker pid=73953) return self._deserialize_msgpack_data(data, metadata_fields)
(RolloutWorker pid=73953) File “/usr/local/lib/python3.10/site-packages/ray/_private/”, line 225, in _deserialize_msgpack_data
(RolloutWorker pid=73953) python_objects = self._deserialize_pickle5_data(pickle5_data)
(RolloutWorker pid=73953) File “/usr/local/lib/python3.10/site-packages/ray/_private/”, line 215, in _deserialize_pickle5_data
(RolloutWorker pid=73953) obj = pickle.loads(in_band)
(RolloutWorker pid=73953) File “/usr/local/lib/python3.10/site-packages/torch/”, line 337, in _load_from_bytes
(RolloutWorker pid=73953) return torch.load(io.BytesIO(b))
(RolloutWorker pid=73953) File “/usr/local/lib/python3.10/site-packages/torch/”, line 1028, in load
(RolloutWorker pid=73953) return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)
(RolloutWorker pid=73953) File “/usr/local/lib/python3.10/site-packages/torch/”, line 1256, in _legacy_load
(RolloutWorker pid=73953) result = unpickler.load()
(RolloutWorker pid=73953) File “/usr/local/lib/python3.10/site-packages/torch/”, line 1193, in persistent_load
(RolloutWorker pid=73953) wrap_storage=restore_location(obj, location),
(RolloutWorker pid=73953) File “/usr/local/lib/python3.10/site-packages/torch/”, line 381, in default_restore_location
(RolloutWorker pid=73953) result = fn(storage, location)
(RolloutWorker pid=73953) File “/usr/local/lib/python3.10/site-packages/torch/”, line 274, in _cuda_deserialize
(RolloutWorker pid=73953) device = validate_cuda_device(location)
(RolloutWorker pid=73953) File “/usr/local/lib/python3.10/site-packages/torch/”, line 258, in validate_cuda_device
(RolloutWorker pid=73953) raise RuntimeError('Attempting to deserialize object on a CUDA ’
(RolloutWorker pid=73953) RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location=torch.device(‘cpu’) to map your storages to the CPU.

This is the snippet for the tuner

tuner = tune.Tuner(
    trainable=add_mixins(config_module.algorithm, [SerializableTorchTrainerMixin]),
        stop={"timesteps_total": config_module.total_timestep},
        callbacks=[TBXLoggerCallback(), CSVLoggerCallback(), JsonLoggerCallback()],
results =

and the scheduler:

scheduler_obj = sched.PopulationBasedTraining(
time_attr=config_module.time_attr_pbt, # name of the metric used to perform a new resampling

It’s kind of tough to properly debug this issue since I cant put break points here and there and this error does not occur when num_gpus is set to 0. So if you have any hint on what is going on and how I can solve it, I would really appreciate it!
Using py3.10, ray 2.9.0 and torch with cuda 118.


I did a bit of trial and error with the parameters but I still struggle to find a final solution.
Nice thing is that the above cuda error seems to go whenever num_gpus_per_worker is set to “some” value. However, I still struggle to clearly see through some elements, which I list here:

  1. why do I need to setup num_gpus_per_worker? I thought only the learner should have gpu allocated (I am not in the case remote workers need gpu).
  2. Why do I get that cuda error if num_gpus_per_worker = 0?
  3. According to the definition, multiple learners can be set in parallel with PBT. What I see is the # of parallel learnings I can do is specified by num_samples, and the effective number of used cpus is num_cpus = num_workers * num_samples, while the effective number of used gpus is num_gpus * num_samples…am I right?
  4. If I specify “too many” num_samples or num_workers, I get cuda out of memory error since (I believe) the # of remote workers is getting to a large value and the model is loaded on gpu also for them - again, am I right in this consideration?

I think I could solve this issue by not assigning gpu to workers, but this currently throws me the cuda error originally specified above.

I will keep digging in and seek a solution, but if someone could share some more insight that would be extremely helpful!