Daniel Han

I'll investigate your new issue!

Unfortunately Kaggle has 20GB of disk space - I would save and upload LoRA / QLoRA adapters, then load them in Google Colab to convert to 16bit

Is this a Mistral model?

@hannesfant Oh is the output correct? They're just warnings

Oh weighting is possible, but you'll need to add a custom cross entropy loss function ie via removing the LM Head, and putting a custom one

You need to use autocasting ie ```python with torch.cuda.amp.autocast(dtype = torch.bfloat16): model(...) ```

Oh ye sadly Windows is an issue - I'm working on something to make Triton an optional dependency

Actually there is - in theory one can install Triton via windows - maybe try https://github.com/unslothai/unsloth/issues/210

Hmm I might make a Windows tutorial in the near future if that helps

How about ```bash sudo apt-get install python3.9-dev sudo apt install libpython3.9-dev ``` Also you need to link it so `ldconfig`