Can xgboost, lightgbm, and pandas be made into optional dependencies?

[Chris Lo (topher-lo)]

Hi flaml community and maintainers! First of all, fantastic library and incredible algorithms. My team is using flaml.tune to run our custom hyperparameter tuning workflows. It's economical, easy-to-use, fast, and powerful: all 💯 features for production.

There is is one problem causing us headaches though: our Docker images with flaml installed are too big. The culprit? xgboost and lightgbm: we aren't using these models in our ML workflow. Having to install these two dependencies just to use flaml.tune is really slowing down our development cycles. Moreover, we are paying for those extra unused MiBs to ECR.

Please correct me if I'm mistaken, but I don't think lightgbm and xgboost are necessary for flaml.tune?

P.S.
Perhaps pandas can also be made into an optional dependency? I know many serious data teams that have pure numpy ML workflows in production: pandas is just unnecessary bloat. Once again, I don't think pandas is required for flaml.tune?