deps: update peft to install from pypi

[anhuong]

relates to: #263

[anhuong]

I tested this by running the tox unit tests and by running the Getting_Started jupyter notebook which ran successfully through the training. I haven't worked with caikit-nlp too much, is there another way that this should be tested? Also, it looks like that on the next release tag it will trigger the pypi push - does this need to be tested?

[gkumbhat]

@anhuong thanks for fixing the package here @anhuong . We would need to do an end to end training test with this, to make sure the training goes through smoothly for seq2seq and causal-lm models and also we are able to do inferencing on the tuned prompts. 🤔

[dtrifiro]

How about just pinning peft==0.6.0 and letting dependabot submit PRs for updates?

[gkumbhat]

LGTM

[gkumbhat]

@anhuong actually lets go with @dtrifiro 's idea of pinning PEFT. That would be a more resilient way of handling for this case.

[gkumbhat]

Lets go with @dtrifiro's suggestion of pinning peft, that would be a bit more resilient to sudden or surprise breakage

going to unblock due to maintainer availability, since requested changes were addressed

[evaline-ju]

Verified training ability and inference after deployment (PT, MPT on causal lm [bloom] and seq2seq [flan-t5-xl])

Edit: update after revert of HF trainer in 0.4.0 - I verified that results were consistent on inference for the same models across version changes

[evaline-ju]

LGTM