Transformers: AutoModel.from_pretrained with torchscript flag raises a TypeError: __init__() got an unexpected keyword argument 'torchscript'

[jonsnowseven]

🐛 Bug

Information

Model I am using: BertModel and AutoModel

Language I am using the model on: English

To reproduce

Steps to reproduce the behavior:

from transformers.modeling_auto import AutoModel
from transformers.modeling_bert import BertModel

bert_model = BertModel.from_pretrained('bert-base-uncased', torchscript=True)
bert_model = AutoModel.from_pretrained('bert-base-uncased', torchscript=True)

Behaviour

bert_model = AutoModel.from_pretrained('bert-base-uncased', torchscript=True) raises a
TypeError: __init__() got an unexpected keyword argument 'torchscript'

Expected behaviour

Successfully create a BertModel object using AutoModel class.

Environment info

• transformers version: 2.9.0
• Platform: Darwin Kernel Version 19.4.0: Wed Mar 4 22:28:40 PST 2020; root:xnu-6153.101.6~15/RELEASE_X86_64
• Python version: 3.6.10
• PyTorch version: 1.3.1
• Tensorflow version: Not applicable
• Using GPU in script?: Not applicable
• Using distributed or parallel set-up in script?: Not applicable

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[benschreiber]

This issue still exists and is relevant to me. Can we get an official response?

[jonsnowseven]

Yes, I would also like an official answer, please.

[benschreiber]

It appears that AutoConfig accepts a torchscript keyword parameter. The AutoConfig object can then be passed as the config keyword parameter to AutoModel. Hope this workaround helps @jonsnowseven

[LysandreJik]

Hi! This was fixed by https://github.com/huggingface/transformers/pull/5665. Could you try to install from source and try again?