Hello,

In the file distributed.py, some errors can be raised suggesting the use of --distributed_launch=True which is misleading because it is not what the user is supposed to do. According to the official documentation of SpeechBrain, if one wants to use DDP then he needs to use --distributed_launch instead of --distributed_launch=True -> https://speechbrain.readthedocs.io/en/v0.5.8/multigpu.html

Moreover, --distributed_launch=True will not work. I tried to do a little example on my GPUs cluster where I'm training an LSTM on the CommonVoice French recipe of SpeechBrain using 2x v100s:

with --distributed_launch=True I got an error: torch.distributed.elastic.multiprocessing.errors.ChildFailedError ......
whearas --distributed_launch do everything that is expected, i.e train the LSTM.

I got the same return from a colleague who tried to use DDP and was stuck due to the errors thrown that were suggesting to set to True.

P.S.: I also need to change every comment in the recipes that suggests that. E.g. CommonVoice ASR seq2seq in train.py -> # If distributed_launch=True but I'm waiting for an official reviewer to confirm the problem.

Best,