The rsm hyperparameter in CatBoost is known to speed up training by using a random subset of features. It’s commonly used during training on CPU but is currently not supported for GPU training, likely because GPU acceleration already provides significant speed improvements.

I am curious about two specific aspects:

• Are there plans to implement the rsm hyperparameter for GPU training in CatBoost in the future?
• Has CatBoost been evaluated to determine the impact of the rsm parameter on model generalization (i.e. overfitting mitigation)?

I have reviewed prior discussions on this topic but could not find clear answers regarding potential deployment plans or the testing of its effects. Any clarification or insights would be greatly appreciated.

Thanks!