fix: add fms-recommender wrapper

[YashasviChaurasia]

Description of the change

The PR adds tuning-config-recommender functionality to fms-hf-tuning simplifying the fine-tuning experience for the end user.

This contributes fms-recommender.py script which simplifies the tuning experience for the user and helps them avoid minor/major pitfalls which can hinder their tuning experience..
With the help of tuning-config-recommender now we can correct erroneous launch commands or get the recommender to provide suggestions for various flags used while fine-tuning.

The fms-recommender.py script has a --preview flag which would preview the suggested command and not execute it..
Note: by default without the --preview flag the suggested command is executed !

Following is a sample of running the command in preview mode:

python fms-recommender.py --preview tuning.sft_trainer \
--model_name_or_path ibm-granite/granite-4.0-h-350m \
--training_data_path tatsu-lab/alpaca \
--tuning_strategy full

Preview command:

Generates a valid data_config file with the datasets, chat-template, data_handlers, etc.

Generates a valid accelerate_config

Note:

Currently the tuning-config-recommender is undergoing review cycles before its first release and hence the pyproject.yaml is not linked to a release version..
for now the pyproject would install via main branch which might not work the same way as shown above due to pending essential PR's for FMSAdapter at : PR's are undergoing review!

The above commands can be replicated via installing recommender from the main branch

Related issue number

How to verify the PR

Was the PR tested

• I have added >=1 unit test(s) for every new method I have added.
• I have ensured all unit tests pass

[github-actions]

Thanks for making a pull request! 😃
One of the maintainers will review and advise on the next steps.

[YashasviChaurasia]

updated pyproject to use main branch until first release.
major changes were merged into main for tuning-config-recommender.

[dushyantbehl]

Do we not have a package for the recommender?

without pypi this could not be included in the release

[YashasviChaurasia]

yes, we do not have a release for the recommender yet!
we can skip this PR for the current minor release if not urgent and required.

[dushyantbehl]

The placement of this wrapper inside tuning is a bit unclear.

Do we maintain separate packages with main inside the same folder?
Can we not place this outside?

[dushyantbehl]

cc @ashokponkumar any suggestions

[YashasviChaurasia]

My reasoning was to keep it at the same level as sft_trainer.py so as the user can simply switch between recommender wrapper or directly use the sft_trainer.py itself..
The goal was to eventually abstract out the usage of sft_trainer.py inside the wrapper so with time the user would only use the wrapper; the wrapper would eventually toggle between recommendation mode vs no-recommendation mode in future updates