Change8

v3.0.3

📦 xgboostView on GitHub →
4 features🐛 6 fixes🔧 6 symbols

Summary

XGBoost 3.0.3 introduces new APIs, adds Rapids 25.06 and Win‑ARM64 wheel support, and includes several bug fixes for GPU evaluation and metrics.

Migration Steps

  1. Verify downloaded artifacts using SHA256 checksums as shown in the release notes.\nsh\necho \"<hash> <artifact>\" | shasum -a 256 --check\n

✨ New Features

  • Added ``setNumEarlyStoppingRounds`` API to JVM packages (#11571).
  • Added support for Rapids 25.06 (#11504).
  • Added ``enable_categorical`` parameter to sklearn ``.apply`` method (#11550).
  • Added support for building XGBoost wheels on Win-ARM64 (#11572, #11597, #11559).

🐛 Bug Fixes

  • Fixed NDCG metric handling for non‑exponential gain (#11534).
  • Fixed RMSLE implementation to avoid using mean intercept (#11588).
  • Prevented implicit synchronization during GPU evaluation (#11542).
  • Removed unnecessary CUDA check in the array interface handler (#11386).
  • Fixed GPU histogram check (#11574).
  • Updated xgboost.testing to be compatible with scikit-learn 1.7 (#11502).

Affected Symbols