How to build onnx quantize fp8 Flux model?

[vuongminh1907]

I have a question that How Black Forest Lab can export fp8 onnx with their checkpoint like this https://huggingface.co/black-forest-labs/FLUX.1-dev-onnx.

I want to build an Onnx fp8 model from my repo vuongminhkhoi4/Flux_HyperSD . Can you take me a guide or give me code for build this? Thanks a lot!

[vuongminh1907]

cc @a-r-r-o-w @asomoza , can you help me

[a-r-r-o-w]

I'm not very familiar with ONNX so maybe I'm pointing you in the wrong direction, but please check out:

• https://docs.pytorch.org/tutorials/beginner/onnx/export_simple_model_to_onnx_tutorial.html
• https://docs.pytorch.org/docs/stable/onnx.html

Diffusers model implementations are just simple pytorch models, so the export should be possible following the guide. There might be a case where some part of our implementation is not export-compatible and need to be modified. In case you run into errors, please let us know :)