auto_select_gpus cannot handle being passed accelerator="gpu", devices=devices

[daniellepintz]

🐛 Bug

I am trying to update test_trainer_with_gpus_options_combination_at_available_gpus_env in #12589 in preparation for #11040, but it is failing with the following stack trace:

tests/trainer/properties/test_auto_gpu_select.py:41:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
pytorch_lightning/utilities/argparse.py:339: in insert_env_defaults
    return fn(self, **kwargs)
pytorch_lightning/trainer/trainer.py:486: in __init__
    self._accelerator_connector = AcceleratorConnector(
pytorch_lightning/trainer/connectors/accelerator_connector.py:194: in __init__
    self._set_parallel_devices_and_init_accelerator()
pytorch_lightning/trainer/connectors/accelerator_connector.py:512: in _set_parallel_devices_and_init_accelerator
    self._parallel_devices = self.accelerator.get_parallel_devices(self._devices_flag)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

devices = None

    @staticmethod
    def get_parallel_devices(devices: List[int]) -> List[torch.device]:
        """Gets parallel devices for the Accelerator."""
>       return [torch.device("cuda", i) for i in devices]
E       TypeError: 'NoneType' object is not iterable

pytorch_lightning/accelerators/gpu.py:82: TypeError

cc @justusschock @kaushikb11 @awaelchli @akihironitta @rohitgr7

[awaelchli]

Was fixed in #12608, just not properly linked.