Closed belerico closed 3 weeks ago
Wow, this is a huge PR! Thanks for working on that! Starting with some high-level feedback, is LongLora always enabled by default? I think we should probably make this optional. Similar to Galore for example. I.e., adding
use_longlora: bool = False,
"""Whether to enable LongLoRA."""
What do you think?
HI @rasbt, I'm closing this because I've seen that there are some unrelated histories containing also commits regarding the removal of the last layers
: my bad because i've put everything together for some experiments I'm running. I'll close this and open a new branch with the history fixed.
Sry about this
I think it might have been possible to rebase, but all good. Next time, here's one approach to do this (thanks to @awaelchli who explained this to me):
# update main
git checkout main
git pull origin main
# update target
git checkout branch-to-rebase
git pull origin branch-to-merge
# make backup
git checkout -b backup-branch-to-merge
# start rebasing
git checkout branch-to-merge
git rebase main branch-to-merge --interactive
# remove overlap
# delete the ones that are already in main
# save and quit
# solve merge conflict
git add file
git rebase --continue
git push -f
This PR introduces LongLora as in #1237 for both the LoRa and full fine-tuning, while also enabling it during generation.