@echarlaix it seems that the TFLite CI is necessary to run to see some failures 🥲
I enabled it for now to avoid any problems, it's fully distributed so shouldn't add an enormous load/wait time.

The core problem seems to be the fact that the tflite exporter uses different APIs (maybe just old) than the onnx exporter (like taking input shapes as input during construction of config).

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

Thanks a lot for checking this @IlyasMoutawwakil ! Looks like optimum TFLite export usage is pretty low, wondering if we should either deprecate if completely in v2 or move everything to optimum-tflite, what do you think?

@echarlaix with the deprecation/removal of TensorFlow in Transformers, I think it only makes sense to deprecate/remove it from optimum, onnx format is imo the universal format for model export from which other frameworks should be able to build their representation (they usually can !)

for example https://github.com/MPolaris/onnx2tflite