Improve the compatibility dealing with large ONNX proto in ORTOptimizer and ORTQuantizer

[JingyaHuang]

What does this PR do?

• Add all_tensors_to_one_file option to ORTOptimizer when exporting the ONNX model.

P.S. only the ONNXModel of optimization has added the option, not quantization.

• Refactoring the compatibility of ORTOptimizer and ORTQuantizer in cases of large ONNX proto(export path / load ONNX)
• Update the optimization and quantization examples with the new option.

Fixes #222

Related to onnx/onnx#4394

P.S.

• Currently the Optimizer exports the original ONNX to the same folder as the optimized ONNX model, ORTOptimizer would better cache the original model as theORTQuantizer, this PR will not target the issue, it could be done either in the refactorization PR ORT optimizer refactorization #294 or in a new PR.
• onnx offers options: save_as_external_data, all_tensors_to_one_file, size_threshold(1024), convert_attribute(false), and the last two are not supported to be self-defined in onnxruntime yet. But in general, the first two are the most effective.
• Currently onnxruntime export external files and model proto after quantization to different directory and this might need to be fixed from their side. Potential updates on ONNXModel.save_model_to_file of quantization and optimization microsoft/onnxruntime#12576

[HuggingFaceDocBuilderDev]

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

[regisss]

Thanks for this PR @JingyaHuang!!
I just have one comment which may be unrelated to this PR.

[regisss]

LGTM, thanks Jingya!!

[JingyaHuang]

Thanks for reviewing, I will put the arguments in ORTModel as suggested and make necessary changes in the examples.

[echarlaix]

LGTM, thanks @JingyaHuang !

[JingyaHuang]

Moved config(model and ORT) saving after saving the model, otherwise will get the following error as configs are saved in the folder:

RuntimeError: Output directory (results/albert_largev1_squad2) for external data is not empty.

While running the template example, I got the error message in the evaluation phase(no problem with the optimization, and models with external files are properly saved):

Traceback (most recent call last):
  File "run_qa.py", line 540, in <module>
    main()
  File "run_qa.py", line 492, in main
    metrics = compute_metrics(predictions)
  File "run_qa.py", line 460, in compute_metrics
    return metric.compute(predictions=p.predictions, references=p.label_ids)
  File "/usr/local/lib/python3.8/dist-packages/evaluate/module.py", line 444, in compute
    output = self._compute(**inputs, **compute_kwargs)
  File "/root/.cache/huggingface/modules/evaluate_modules/metrics/evaluate-metric--squad/b4e2dbca455821c7367faa26712f378254b69040ebaab90b64bdeb465e4a304d/squad.py", line 110, in _compute
    score = compute_score(dataset=dataset, predictions=pred_dict)
  File "/root/.cache/huggingface/modules/evaluate_modules/metrics/evaluate-metric--squad/b4e2dbca455821c7367faa26712f378254b69040ebaab90b64bdeb465e4a304d/compute_score.py", line 67, in compute_score
    exact_match += metric_max_over_ground_truths(exact_match_score, prediction, ground_truths)
  File "/root/.cache/huggingface/modules/evaluate_modules/metrics/evaluate-metric--squad/b4e2dbca455821c7367faa26712f378254b69040ebaab90b64bdeb465e4a304d/compute_score.py", line 52, in metric_max_over_ground_truths
    return max(scores_for_ground_truths)
ValueError: max() arg is an empty sequence

Seems like a bug with the metric, an update might need to be done with the examples. I can take it over if you don't have bandwidth @echarlaix.