Benjamin Bossan

This is strange, there are sklearn versions on pypi that satisfy this version requirement: https://pypi.org/project/scikit-learn/#history I tried `pip install dstoolbox==0.9.1` (Python 3.7) and it worked, installing `scikit-learn-0.22.2.post1`. Are you sure...

Sorry, I don't understand your issue. Could you please explain it in more detail, what you want to achieve and how? Ideally show the code that leads to good or...

I see. I don't really have any expertise in whisper and how the initial prompt affects the outcome. But my best guess is that yes, if you want to use...

Hmm, hard to say and I can't easily try to reproduce this. Do you already see strange behavior after loading the model, before starting training? If you try without PEFT,...

Can you try to run this additional snippet: ```python model = get_peft_model(...) # convert all peft parameters to float32 for param in model.parameters(): if param.requires_grad: param.data = param.data.float() ```

> when merging the weights... I suppose currently it will work for `nn.Params` but now it also needs to do the same for the buffers... how is that handled? Hmm,...

Oh, now I wonder if there isn't a different solution: Ask users to add the batch norm layers to `modules_to_save`. I haven't thought this through completely, but this may solve...

Just tested the `modules_to_save` solution: When using this config: `LoraConfig(target_modules=["convolution"], modules_to_save=["classifier", "normalization"])`, the test passes even without changing the code to explicitly add buffers to the `state_dict`.

@pacman100 I removed the new functionality and instead changed the test to add batch norm layers to `modules_to_save` and the tests pass. I also added a troubleshooting section about this.

You're right, the comment doesn't match the code. After a quick glance at the LoRA paper, I don't see an explicit mention of how LoRA should be initialized for embedding...