ONNXOptimizer: ValueError: Message onnx.ModelProto exceeds maximum protobuf size of 2GB: 2215158499

@yufenglee I'm not sure this issue is fixed (tried using the latest onnxruntime-1.12.0 GPU). While disabling shape inference works for quantization, optimization is still broken for saving models such as albert-xxl. The issue seems to be that write_external_data_tensors isn't actually respecting the condition when it's set to true:

Traceback (most recent call last):
File "examples/onnxruntime/optimization/question-answering/run_qa.py", line 525, in
main()
File "examples/onnxruntime/optimization/question-answering/run_qa.py", line 311, in main
optimizer.export(
File "/home/mroyzen/train_files/optimum/optimum/onnxruntime/optimization.py", line 149, in export
optimizer.save_model_to_file(onnx_optimized_model_output_path, use_external_data_format=True)
File "/opt/conda/lib/python3.8/site-packages/onnxruntime/transformers/models/gpt2/../../onnx_model.py", line 934, in save_model_to_file
save_model(self.model, output_path)
File "/opt/conda/lib/python3.8/site-packages/onnx/init.py", line 202, in save_model
s = _serialize(proto)
File "/opt/conda/lib/python3.8/site-packages/onnx/init.py", line 71, in _serialize
result = proto.SerializeToString()
ValueError: Message onnx.ModelProto exceeds maximum protobuf size of 2GB: 2215158499

Your help would be appreciated. Linking to my issue in Huggingface's Optimum library, but everything seems to be working on their end as evidenced by the stack trace. Also linking the related ONNXQuantizer issue.

Your help would be appreciated. Thanks

Originally posted by @michaelroyzen in https://github.com/microsoft/onnxruntime/issues/11072#issuecomment-1197711143