move check_dummy_inputs_allowed to common export utils

[eaidova]

What does this PR do?

check_dummy_inputs_allowed is helpful function for any optimum that involves pytorch models conversion (e.g. openvino). Current place of it import trigger initialization of full onnx configuration for other backends that wish to reuse it and it may lead to some troubles with resolving dependencies. I suggest to move it to common utilities for avoiding unexpected errors like represented bellow:

Example of logs (issue visible on windows with onnx>1.16.2 and related to onnx/onnx#6267):

  File "C:\Users\chuquanc\openvino\py310\lib\site-packages\optimum\intel\openvino\__init__.py", line 51, in <module>
    from .quantization import OVQuantizer
  File "C:\Users\chuquanc\openvino\py310\lib\site-packages\optimum\intel\openvino\quantization.py", line 47, in <module>
    from optimum.exporters.onnx.convert import check_dummy_inputs_are_allowed
  File "C:\Users\chuquanc\openvino\py310\lib\site-packages\optimum\exporters\__init__.py", line 16, in <module>
    from .tasks import TasksManager  # noqa
  File "C:\Users\chuquanc\openvino\py310\lib\site-packages\optimum\exporters\tasks.py", line 173, in <module>
    class TasksManager:
  File "C:\Users\chuquanc\openvino\py310\lib\site-packages\optimum\exporters\tasks.py", line 338, in TasksManager
    "clip-text-model": supported_tasks_mapping(
  File "C:\Users\chuquanc\openvino\py310\lib\site-packages\optimum\exporters\tasks.py", line 117, in supported_tasks_mapping
    importlib.import_module(f"optimum.exporters.{backend}.model_configs"), config_cls_name
  File "C:\Users\chuquanc\packages\Python310\lib\importlib\__init__.py", line 126, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
  File "C:\Users\chuquanc\openvino\py310\lib\site-packages\optimum\exporters\onnx\model_configs.py", line 23, in <module>
    from ...onnx import merge_decoders
  File "<frozen importlib._bootstrap>", line 1075, in _handle_fromlist
  File "C:\Users\chuquanc\openvino\py310\lib\site-packages\transformers\utils\import_utils.py", line 1766, in __getattr__
    module = self._get_module(self._class_to_module[name])
  File "C:\Users\chuquanc\openvino\py310\lib\site-packages\transformers\utils\import_utils.py", line 1780, in _get_module
    raise RuntimeError(
RuntimeError: Failed to import optimum.onnx.graph_transformations because of the following error (look up to see its traceback):
DLL load failed while importing onnx_cpp2py_export: A dynamic link library (DLL) initialization routine failed.

workaround for this problem also proposed on optimum-intel side huggingface/optimum-intel#1048

[eaidova]

@echarlaix @IlyasMoutawwakil could you please review?

[IlyasMoutawwakil]

LGTM ! Thanks !

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[echarlaix]

would prefer to keep import there, why moving ?

[eaidova]

merge_decoders imported from graph_transformations submodule that contains onnx import
https://github.com/huggingface/optimum/blob/main/optimum/onnx/graph_transformations.py#L19

as this file heavily depends on onnx functionality, while merge decoders used only for specific configs postprocessing triggered only if specific config behaviour is selected, I think it may be better to allow use other functional from this module without necessarily to use onnx import.

As I said there is case, when import onnx crashed on windows. This file contains basic functionality for configs that we also reuse for enabling new models in optimum-intel moving this import allow us continue usage of base config classes even if onnx broken