Add support for Falcon model to export to ONNX

[mindest]

Feature request

Add support so that we can export Falcon model to ONNX format using optimum-cli, like

optimum-cli export onnx -m tiiuae/falcon-7b falcon_7b --trust-remote-code

Motivation

Wanted to run Falcon model using ONNXRuntime. Currently error will be raised when exporting Falcon model to ONNX, using the above command:
KeyError: "refinedwebmodel is not supported yet. Only {'electra', 'mt5', ...} are supported. If you want to support refinedwebmodel please propose a PR or open up an issue."

Your contribution

Support may be dependent on updates in transformers, like huggingface/transformers#24523.

[fxmarty]

Hi @mindest , yes the easiest way would be first a transformers support we can base ourselves on. An alternative is to indeed use the custom modeling from tiiuae/falcon-7b & to use custom ONNX configs for the export (available only on main for now, not in a release: https://huggingface.co/docs/optimum/main/en/exporters/onnx/usage_guides/export_a_model#customize-the-export-of-transformers-models-with-custom-modeling)

[fxmarty]

@mindest Actually huggingface/transformers#24523 has been merged, so just waiting on a release from transformers :)

[Johnno1011]

Yo, just curious, why do we have to wait for a proper release (IE: pip etc) before this can be resolved? I was able to build from source and it works just fine.
Thanks.

[fxmarty]

Our CI relies on transformers latest release.

[Pradipve2011]

@Johnno1011

Hey John,

I am looking for solution on this. Can you please share how do you able to convert it?

[WilliamTambellini]

up

[fxmarty]

Hi, the support is added in #1391