please add option to load fine-tuned file to CPU if trained on GPU

[bsugerman]

I fine-tuned the pytorch_model.bin on a GPU machine (google cloud) but need to use it on my home computer (no GPU). When I tried to open it using model = BertForMaskedLM.from_pretrained(bert_version) I got the following error:

 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='cpu'
 to map your storages to the CPU.

Perhaps you can add an option into from_pretrained() such as cpu=True which will then call

torch.load(weights_path, map_location=lambda storage, location: 'cpu')

[thomwolf]

Indeed, this will be in the next release, thanks!

[ahmedbahaaeldin]

is there is a change according the CPU issue ?

[thomwolf]

Should be fixed now. Do you still have an error?

[ahmedbahaaeldin]

i am loading the model and i dont know how to load on CPU , gives me "model.to" to is not defined. Can you tell me how to send the model to run on CPU if trained on GPU.

[CBvanYperen]

Would love some explanation on how to do this as well!

[Zhen-hao]

I have the same question. how to load a model trained on GPU to CPU?

[ariellasmo]

I am watching this as well.

[FullDataAlchemist]

I have the same question. Is there any option for using CPU-only?

[davidsbatista]

watching this as well

[iliaNecrov]

same