Grace Blackwell 上使用 Spark Connect ML 扩展 XGBoost.pdf

1、Forward-looking StatementThis presentation has been prepared for informational purposes only.The information set forth herein does not purport to be complete or contain all relevant information.Statements contained herein are made as of the date of this presentation unless stated otherwise.This pres

2、entation and the accompanying oral commentary may contain forward-looking statements.In some cases,forward-looking statements can be identified by terms such as“may”,“will”,“should”,“expects”,“plans”,“anticipates”,“could”,“intends”,“projects”,“believes”,“estimates”,“predicts”,or“continue”,or the neg

3、ative of these words or other similar terms or expressions that concern Databricks expectations,strategy,plans,or intentions.Forward-looking statements are based on information available at the time those statements are made and are inherently subject to risks and uncertainties that could cause actu

4、al results to differ materially from those expressed in or suggested by the forward-looking statements.Forward-looking statements should not be read as a guarantee of future performance or outcomes.Except as required by law,Databricks does not undertake any obligation to publicly update or revise an

5、y forward-looking statement,whether as a result of new information,future developments or otherwise.2Scaling XGBoost With Spark Connect ML on Grace BlackwellJiaming Yuan,Bobby Wang4/12/2025Why Are We Here?Main theme of the talkScalingScalingScalingWe are here to scaleXGBoost trainingAgendaIntroducti

6、onExternal Memory Model TrainingScaling with NVLink-C2CJiaming YuanSpark Connect MLRun XGBoost JVM package over spark connectBobby WangPart IIntroduction to the Out-of-Core XGBoostJiaming YuanIntroductionXGBoostIntroductionGradient boosting decision treesSupports multiple target typesRegressionClass