search

microsoft / onnxconverter-common

Common utilities for ONNX converters
MIT License
250 stars 66 forks source link

Max supported opset is very old #296

Open addisonklinke opened 3 months ago

addisonklinke commented 3 months ago

DEFAULT_OPSET_NUMBER is currently 15 and was last updated in Nov 2021. This corresponds to a max of onnx==1.10.2 from the official versioning table which is 6 minor versions behind the latest 1.16.0. Additionally, 1.10.2 only has wheels for Python <=3.9 which is EOL Oct 2025 and makes using a more modern env difficult

What are the limitations in upgrading this / why is it lagging so far behind the onnx releases?

MaanavD commented 3 months ago

Hey, I'm assuming you're referencing the torch.onnx.export() api. If you can try torch.onnx.export(...,dynamo=True) you might have a better outcome!

addisonklinke commented 3 months ago

Actually my use-case is in PySpark and not torch. From my understanding, DEFAULT_OPSET_NUMBER is internal to onnxconverter-common so any 3rd party library which uses it for conversion would be limited to that opset

addisonklinke commented 2 months ago

@MaanavD any direction for how we could bump the opset version?

MaanavD commented 2 months ago

@addisonklinke going to add @gramalingam to this thread, I think he'd know best!

gramalingam commented 2 months ago

@xadupre knows more about these converters.