Feature Request: Native MultiOutput Support on GPU (CatBoost-style)

[HamdiBarkous]

Training a multioutput regression model on GPU with XGBoost is significantly slower (~6x) compared to CatBoost.

Setup:

• 50k samples, 500 features, 15 targets
• Default GPU params

Observation:
XGBoost appears to train one model per target, while CatBoost handles multioutput natively, leading to major performance gains.

Can you add native multioutput support on GPU for better performance, similar to CatBoost.

[hcho3]

XGBoost provides support for multi-output trees, starting in version 2.0. Check out the tutorial here.

Note that support for muti-output trees is still experimental. Bug reports are appreciated.

[trivialfis]

Yes, at the moment XGBoost fits one model per target. That could be part of the reason and we are working on it. (vector leaf)

[trivialfis]

Closing in favor of #9043 .