AttributeError: module 'typing' has no attribute '_ClassVar'

[HatterTheMad]

Environment info

• transformers version: 3.5.1
• Platform: Linux
• Python version: 3.8
• PyTorch version (GPU?): 1.7.0+cpu
• Using GPU in script?: No
• Using distributed or parallel set-up in script?: Yes

Who can help @sgugger .... @LysandreJik ... mb?

Information

Model I am using: Distilbert

The problem arises when using:
Just this:

from transformers import AutoTokenizer, AutoModelForQuestionAnswering

The tasks I am working on is:

• my own task or dataset:
  I am using default Distilbert for my flask API

To reproduce

Steps to reproduce the behavior:
That's a big part of the question. It works just fine on my local machine, but gives this error when run on my AWS server.

from flask import Flask, request, jsonify
from flask import Flask
from flask_restful import Api, Resource
from transformers import AutoTokenizer, AutoModelForQuestionAnswering
from transformers.pipelines import pipeline

tokenizer = AutoTokenizer.from_pretrained('./Dis_Save/')
model = AutoModelForQuestionAnswering.from_pretrained('./Dis_Save/')
nlp_qa = pipeline('question-answering', tokenizer=tokenizer,model=model)

app = Flask(__name__)

@app.route('/api/QandA', methods=['GET', 'POST'])

def QandA():
    content = request.json
    print(content['userMessages'])
    X = nlp_qa(context=content['userMessages'], question=content['question'])
    return(jsonify({"answer":X["answer"], "score":X["score"]}))

if __name__ == "__main__":
	app.run(debug=True)

#This is all the code that I have. Here is the full error that I get:

  File "access/main.py", line 4, in <module>
    from transformers import AutoTokenizer, AutoModelForQuestionAnswering
  File "/home/ubuntu/access/transformers/__init__.py", line 22, in <module>
    from .integrations import (  # isort:skip
  File "/home/ubuntu/access/transformers/integrations.py", line 82, in <module>
    from .trainer_callback import TrainerCallback  # noqa: E402
  File "/home/ubuntu/access/transformers/trainer_callback.py", line 27, in <module>
    from .training_args import TrainingArguments
  File "/home/ubuntu/access/transformers/training_args.py", line 36, in <module>
    class TrainingArguments:
  File "/home/ubuntu/access/dataclasses.py", line 958, in dataclass
    return wrap(_cls)
  File "/home/ubuntu/access/dataclasses.py", line 950, in wrap
    return _process_class(cls, init, repr, eq, order, unsafe_hash, frozen)
  File "/home/ubuntu/access/dataclasses.py", line 800, in _process_class
    cls_fields = [_get_field(cls, name, type)
  File "/home/ubuntu/access/dataclasses.py", line 800, in <listcomp>
    cls_fields = [_get_field(cls, name, type)
  File "/home/ubuntu/access/dataclasses.py", line 659, in _get_field
    if (_is_classvar(a_type, typing)
  File "/home/ubuntu/access/dataclasses.py", line 550, in _is_classvar
    return type(a_type) is typing._ClassVar
AttributeError: module 'typing' has no attribute '_ClassVar'

[LysandreJik]

This is weird! Looking at the stacktrace, I have a few questions:

• Are you running a file in access/main.py?
• If yes, is there a dataclasses.py file in that access folder?

It seems far-fetched but if that's the case, it might be possible that this file is interfering with the dataclasses module.

[attardi]

It seems related to an incompatibility of python > 3.6 with package dataclasses, as explained here:

    https://github.com/google/flax/pull/270

[LysandreJik]

Yes, but we only install dataclasses on Python versions that are inferior to 3.7: https://github.com/huggingface/transformers/blob/master/setup.py#L137

[nsaef]

I encountered the same problem with the following setup:

• transformers version: 3.5.1
• Platform: Linux
• Python version: 3.8
• PyTorch version (GPU?): 1.7.0+cpu
• Using GPU in script?: No
• Using distributed or parallel set-up in script?: No

Local execution works fine, but when running the code on Google App Engine (Standard Environment), it fails with error AttributeError: module 'typing' has no attribute '_ClassVar'. There is no file called dataclasses.py anywhere in the project,

Stacktrace:

File "/srv/application/stance_pred/bert_inference.py", line 8, in <module>      from transformers import DistilBertForSequenceClassification, DistilBertTokenizer, DistilBertConfig
File "/layers/google.python.pip/pip/transformers/__init__.py", line 22, in <module>      from .integrations import (  # isort:skip
File "/layers/google.python.pip/pip/transformers/integrations.py", line 82, in <module>      from .trainer_callback import TrainerCallback  # noqa: E402
File "/layers/google.python.pip/pip/transformers/trainer_callback.py", line 27, in <module>      from .training_args import TrainingArguments
File "/layers/google.python.pip/pip/transformers/training_args.py", line 36, in <module>      class TrainingArguments:
File "/layers/google.python.pip/pip/dataclasses.py", line 958, in dataclass      return wrap(_cls)
File "/layers/google.python.pip/pip/dataclasses.py", line 950, in wrap      return _process_class(cls, init, repr, eq, order, unsafe_hash, frozen)
File "/layers/google.python.pip/pip/dataclasses.py", line 800, in _process_class      cls_fields = [_get_field(cls, name, type)
File "/layers/google.python.pip/pip/dataclasses.py", line 800, in <listcomp>      cls_fields = [_get_field(cls, name, type)
File "/layers/google.python.pip/pip/dataclasses.py", line 659, in _get_field      if (_is_classvar(a_type, typing)
File "/layers/google.python.pip/pip/dataclasses.py", line 550, in _is_classvar      return type(a_type) is typing._ClassVar  AttributeError: module 'typing' has no attribute '_ClassVar'

[thesby]

Any solution?

[LysandreJik]

Could one of you post the result of pip list in the environment where that is failing? Or even better paste the result of pip freeze, alongside a few lines of code that reproduce the issue.

Thank you!

[thesby]

I solved this problem by removing dataclasses*

[HatterTheMad]

I have solved this problem by downgrading to Python version: 3.6

Thanks @attardi

[stas00]

note that fairseq 0.10.1 requires dataclasses even for py>3.7 where it's built-in and HF Trainer breaks when dataclasses are installed for these versions. So if some project pulls in fairseq which will force the install of dataclasses HF Trainer will break. Probably need to ask fairseq to fix their dependencies.

Until then @thesby's solution is the easiest one.

pip uninstall dataclasses -y

[vicesodan]

I don't know why, but for me changing the version of Python worked (from 3.9.12 to 3.9.5.)

[loretoparisi]

note that fairseq 0.10.1 requires dataclasses even for py>3.7 where it's built-in and HF Trainer breaks when dataclasses are installed for these versions. So if some project pulls in fairseq which will force the install of dataclasses HF Trainer will break. Probably need to ask fairseq to fix their dependencies.

Until then @thesby's solution is the easiest one.

pip uninstall dataclasses -y

In my case pip command itself was broken so I had to do by hand:

rm -rf lib/dataclasses-0.6.dist-info
rm lib/dataclasses.py

in the location where dataclasses has been installed!

[wookayin]

Dependency to dataclasses in setup.py should be python 3.6 only:

-'dataclasses'
+'dataclasses; python_version < "3.7"'

or use if sys.version_info >= (3, 7): check to dynamically configure _deps in https://github.com/huggingface/transformers/blob/main/setup.py#L98.

[benj3542]

I met the same error without having dataclasses installed in my environment (python 3.9). I installed dataclasses and uninstalled it again - from this point on the error disappeared.