passing device_map argument breaks the model
#6
by
x75
- opened
Hi, I'm experimenting with MERT, thanks for making & providing it.
I noticed that whenever I am passing a "device_map" argument to AutoModel.from_pretrained the model only outputs nans. This happens regardless of the value of the argument, eg "cuda" or "cpu". When I remove the argument from the call, it works fine.
Initially observed on transformers==4.31.0, now I upgraded to 4.37.1 but its the same.
cheers