wanton marsh
Hello,
I was looking into the notebooks to see how I should load a model that I had previously trained. I notice that in the Qwen3 and LLama3, for example, the FastLanguageModel class was used to load the model. But for Gemma3, the FastModel is used. In fact, if I use the FastLanguageModel class for Gemma3, the default model is being loaded and not the PeftModel. The thing is, I am getting very good results using Gemma3 and terrible with the other models. So I tried to use the FastModel for loading the other models, but this time, when I use the model, a type mismatch error occurs RuntimeError: self and mat2 must have the same dtype, but got Half and Float
How should I choose what class to use? Should I just follow the notebooks? Could it be possible that the Gemma3 model could be this good?
PS: The training setup is exactly the same for all the models.
shadow stirrup