Clarity on the convert.py for a model to ONNX.py.. documentation issue about optimum HOT 2 CLOSED

Hi @gidzr, you should be able just to do pip install optimum and use optimum-cli export onnx --help. More details here: https://huggingface.co/docs/optimum/main/en/exporters/onnx/usage_guides/export_a_model

I believe @xenova relies on Optimum ONNX export but has its own quantization script.

Also, in order to run the optimum/exporters/onnx/convert.py script, do I need to download the full exporters folder, just the onnx folder, or can I just copy-paste the script and run that indepdently?

Normally you don't need to copy anything - if you install through pip you can just run the command above. Of course you would like do to an editable install you can clone the repo and use pip install -e ..

from optimum.

Hi @fxmarty, thanks kindly for that.

Yep re xenova's own export. I'll drop him a line on that. I'm currently running it as a standalone script as xenova's version includes params for --per_channel true --reduce_range true, which made a big difference on the quantisation reliability. I've had the best results with his script.

I'm sure it's on the roadmap, but would be great to see a unified package that incorporates a consistent set of arguments with all the good stuff from optimum and xenova's approaches available a separate pip module or online at https://huggingface.co/spaces/onnx/export would be great. There's definitely few different flavours of opt/quant scripting at the moment, including the in-inference model.save version. Probably a full-time project to consolidate all of these.

from optimum.