h2o-llmstudio
h2o-llmstudio copied to clipboard
Add falcon peft target modules
This PR adds default lora target layers for falcon models. I have excluded MPT models, as they seem to require additional code changes.
@maxjeblick In best case I would envision it that we pre-fill the lora target modules field in settings automatically after switching a backbone. Do you think that's possible?
That's a good suggestion, and certainly improves UX!
Not sure what's the best way to add this, switching backbones would need to trigger an update of ConfigNLPCausalLMTraining parameters. Maybe add an .update method to ConfigProblemBase that can changes parameters based on current cfg and call that method before get_ui_elements. WDYT? It would be a bit similar to nesting, but with the ability to connect different configuration classes.
RLHF may also benefit from such an update function to hide/change certain parameters, cc. @pascal-pfeiffer
Yes, I believe an additional functionality is needed. Would be great if you give it a shot and as you said it might be useful for other stuff.
@psinger The potential UI implementation mentioned above will probably take some time and effort to work correctly. If it's ok with you, I'd merge this PR as is (or without the automatic updates, at least) to fix failing falcon experiments and work on potential UI/UX improvements afterwards.