Benjamin Bossan

> what I observed was version mismatch, which doesn't happen in the production, I know, but I don't think version mismatch should concern the person if the predictions are returned...

> Does predictions change according to versions of sklearn? sweat_smile Do the implementation changes? Without knowing all the details, I think for the vast majority of cases, predictions won't change....

> Should we consider showing the predictions, even if there is a mismatch, and expose the warnings below the widget? Just to note, this part of the error message, `"output":...

> This is not a great UX as it's confusing for users. Even if we show the text of the warning in the widget, this will not be super useful...

I was wondering if we can change the logic here: https://github.com/huggingface/api-inference-community/blob/778fe8446856a392646d11437cb634d979978be0/docker_images/sklearn/app/pipelines/tabular_classification.py#L70-L80 Right now, if there are any warnings, we just treat it as an error. Perhaps we can make an...

Yes, the waiting time is annoying. I think it's good to have the option to choose a specific task. Ideally, there should still be the possibility to run all tasks...

Thanks for the ping. The issue is that PEFT merges the LoRA weights into the `lm_head`, since you added it to `target_modules`. However, the weight of the LM head are...

> I suppose that if with LoRA I avoid targetting LM head but still target embeddings, after the merge_and_unload() the LM head will get mutated again resulting in wrong results...

Thanks for the heads up

#41