I am a long time xgb user but usually I don’t track difference on xgb versions until I ran into bugs today. I tried to fix issues with my GPU so I upgraded drivers and also upgraded xgboost to 1.7.x from 1.3.x (or 1.2.x) by the way. I immediately noticed that the GPU memory usage for same dataset boosted from 3.5 GB to 6.5 GB and the system memory usage was 38 GB. I frequently ran into page fault or illegal memory issues. I could not even complete a successful job which consists a dozen of trainings. For the trainings completed, the time is about 60% of the original trainings. No good solutions found after lengthy web searches. I then downgraded xgboost to 1.3.x and now the first training is in progress. The memory usage is now 2.5GB and system memory usage is 43.x GB.
Is this difference caused by version differences?