Giters

themrzmaster / git-re-basin-pytorch

Git Re-Basin: Merging Models modulo Permutation Symmetries in PyTorch

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Difference between 2 LLM models

NamburiSrinath opened this issue · comments

commented

Hi @themrzmaster,

Thanks for implementing Git-Rebasin in Pytorch

I've a quick question and am wondering whether it's feasible with the current repo:

  1. Suppose I've a model (Llama-1) with weights original_model.pt
  2. Assume, I fine-tuned/modified the model for some use-case and let the weights be modified_model.pt

My high-level question is to understand the difference between these 2 functions (i.e difference between original_model.pt and modified_model.pt!) in a quantitative measure. I am assuming your paper deals with similar stuff (the computed Barrier is basically a quantitative measure which tells the difference between these 2 functions).

Is my understanding correct? If so, can you give some instructions on how your repo can be extended to Huggingface models (Llama, Vicuna, GPT, T5 etc;) I am assuming the logic stays the same!

If not, please provide some insights on how this use-case can be done! I'm assuming Wasserstein, MMD might be the next best bet but would like to try your repository!

P.S: Opened a thread in original repo (samuela/git-re-basin#13 (comment)) can close either one depending on response as the authors have not worked extensively on Pytorch models.

Any help is super appreciated :)
Thanks
Srinath