BUG: SHAP values calculated using CPU differ from SHAP values calculated using GPU for XGBOOST

[madakkmi]

There are large differences between SHAP values calculated using CPU and GPU for XGBoost models with feature_perturbation='interventional' and model_output='log_loss'. The detailed description of the bug including how to reproduce it and the traceback are provided in shap/shap#3655. The issue seems to be originating at the shap implementation of the xgboost resposiory as mentioned in that issue. It would be great if this can be fixed.

[trivialfis]

cc @RAMitchell .