PEFT Implementation

#24
by SebastianBodza - opened

It seems like the PEFT Implementation is not done yet. When trying to train a LORA following error occurs:
TypeError: MPTForCausalLM.forward() got an unexpected keyword argument 'inputs_embeds'

Any Information if or when this will be implemented?

have been looking at this. It looks like this might be a patch:
https://github.com/iwalton3/mpt-lora-patch

However, I didn't really look at the code nor am I sure if there is a related pull request or something else in the works.

The author said it worked but then doesn't seem sure. From the patch author: "The model code specifically says that left padding is not supported, but I forcibly did so and got decent results."

Works perfectly thanks :)

SebastianBodza changed discussion status to closed

@SebastianBodza where to use the patched version.
How to run it. which file to run.
Can you help me

Sign up or log in to comment