Fixes for ORT 1.13.1

[regisss]

What does this PR do?

This PR fixes a couple of errors that appeared with ORT 1.13.1:

• disable_shape_inference is added to OptimizationConfig since it has been added to FusionOptions in ORT (see Add --disable_shape_inference option to optimizer.py microsoft/onnxruntime#12215),
• replace input_qType by activation_qType in quantization.py since this argument name has changed in ORT (see Splitting quantize_tensor and quantize_input microsoft/onnxruntime#12873).

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[echarlaix]

Thanks for taking care of this @regisss !

[echarlaix]

In order for us to support onnxruntime version lower than 1.13.1, could you adapt the quantizer_factory behavior depending on which onnxruntime version is used ?

[regisss]

You're absolutely right, I just added it!

[echarlaix]

Thanks a lot @regisss

[ghost]

Thank you very much @regisss!

Especially since opimization is currently not working for me because of this missing field:

File "/srv/./app/steps/transformers/convert_transformer_to_onnx_step.py", line 43, in convert_transformer_to_onnx_step
    _ = ORTOptimizer.from_pretrained(
        │            └ <classmethod object at 0x7f9e158abf70>
        └ <class 'optimum.onnxruntime.optimization.ORTOptimizer'>

  File "/usr/local/lib/python3.9/site-packages/optimum/onnxruntime/optimization.py", line 128, in optimize
    optimization_options = FusionOptions.parse(optimization_config)
                           │             │     └ OptimizationConfig(optimization_level=1, optimize_for_gpu=False, fp16=False, optimize_with_onnxruntime_only=False, disable_ge...
                           │             └ <staticmethod object at 0x7f9e15633340>
                           └ <class 'onnxruntime.transformers.fusion_options.FusionOptions'>
  File "/usr/local/lib/python3.9/site-packages/onnxruntime/transformers/fusion_options.py", line 64, in parse
    if args.disable_shape_inference:
       └ OptimizationConfig(optimization_level=1, optimize_for_gpu=False, fp16=False, optimize_with_onnxruntime_only=False, disable_ge...

AttributeError: 'OptimizationConfig' object has no attribute 'disable_shape_inference'

Does anybody know a workaround or should I just wait for the patch? Will it be also cherry picked into 1.3.1 or only directly to 1.4.1?

Thanks!

[regisss]

Hi @nmaoez! If you need the fix ASAP, the best is to install Optimum from source as follows:

git clone https://github.com/huggingface/optimum.git
cd optimum/
pip install .[onnxruntime]

A patch is coming very soon but I don't know if this will be integrated to it. Maybe @echarlaix can tell more about this.

[echarlaix]

Hi @nmaoez @regisss, the optimum release 1.4.1 is out

[ghost]

@echarlaix Amazing! Thanks!