ludwig-ai / ludwig

Low-code framework for building custom LLMs, neural networks, and other AI models
http://ludwig.ai
Apache License 2.0
11.19k stars 1.19k forks source link

Remove target_module LoRA mapping for Phi-2 model #3910

Closed arnavgarg1 closed 4 weeks ago

arnavgarg1 commented 10 months ago

The current PEFT LoRA modules map to the wrong target modules for Phi-2 since the architecture was recently updated.

I created a PR on the PEFT repo to update this to new target modules: https://github.com/huggingface/peft/pull/1375

For now, we will fall back to target_modules for Phi-2 that lead to good model performance.