Improve visibility of GPU training in Tabular

Currently users can get various models such as LightGBM, CatBoost, XGBoost and tabular neural networks to train with GPU, however it is not well documented.

The current process to enable GPU for tabular models:

predictor.fit(..., ag_args_fit={'num_gpus': 1})

Note that LightGBM may need a special installation to use GPU.

[Innixma]

So it should be like “ ‘num_gpus’: 8 ”if I have eight gpus right?

Yes, but currently we have not tested multi-gpu and it may only use 1 GPU depending on the model.

[Innixma]

but i got :Training LightGBM with GPU, note that this may negatively impact model quality compared to CPU training. So should i do not add ‘ag_args_fit’?

This message is because model quality is not identical between CPU and GPU, and GPU may have worse results. This is something you would need to verify yourself if GPU model is attaining good results for your needs.