Update XGBoost max supported version to 3.0.2

[sid22669]

Motivation

Fixes #2596

Users are blocked from converting XGBoost 3.x models because _XGBOOST_MAX_VERSION is set to 1.4.2. XGBoost 1.4.2 is incompatible with newer Python versions (3.12+), forcing users into a dependency nightmare.

Modifications

coremltools/_deps/__init__.py:

• Updated _XGBOOST_MAX_VERSION from "1.4.2" to "3.0.2"

Verification

Tested all conversion paths with XGBoost 3.0.2 — the converter is fully compatible:

• XGBRegressor → CoreML spec: OK
• XGBClassifier (binary) → CoreML spec: OK
• XGBClassifier (multi-class) → CoreML spec: OK
• Raw Booster → CoreML spec: OK

The JSON tree dump format (split, split_condition, children, leaf, cover, yes, no, missing) is unchanged between 1.4.2 and 3.0.2. All APIs used by the converter (get_booster(), get_dump(), feature_names, copy(), n_classes_) work identically.

Checklist

• Format code with pre-commit
• Follow coremltools code style

[TobyRoseman]

The tests are still using the old version of XGBoost. See the link I shared in the issue.

[sid22669]

Thanks for the review! I've updated reqs/test.pip to use XGBoost 3.0.2.

While testing, I also found two compatibility issues with XGBoost 3.x:

1. base_score auto-estimation — XGBoost 3.x auto-estimates base_score from training data. The converter was hardcoding 0.5/0.0, causing prediction mismatches. Fixed by reading the actual value from booster.save_config().

2. feature_names type — XGBoost 3.x requires feature_names to be a list. Added a type check before assignment.

Added a test to verify base_score propagation. Will push these changes shortly.

[TobyRoseman]

Thanks for removing the try/except.

CI: https://gitlab.com/coremltools1/coremltools/-/pipelines/2376256024

[JiwaniZakir]

The new XGBoostBaseScoreTest only exercises the regressor path, but the diff also modifies base_prediction for both binary and multiclass classifiers in _tree_ensemble.py (lines ~252–265). The classifier branches—where base_score from the model config is spread across n_classes slots—have no corresponding test coverage, which is worth adding given that the semantics of base_score differ between regression and classification objectives (e.g., log-odds vs. raw margin for binary logistic).

Additionally, load_boston() was deprecated in scikit-learn 1.2 and removed in 1.4; the test class's setUpClass will fail on any modern scikit-learn installation. The existing tests in the file use it too, but introducing new test cases that rely on it compounds the problem rather than addressing it—switching to a synthetic dataset via sklearn.datasets.make_regression would be straightforward and future-proof.

Finally, the config['learner']['learner_model_param']['base_score'] key path is duplicated verbatim in both the converter (_tree_ensemble.py:203) and the test (test_boosted_trees_regression_numeric.py:332). Extracting a small helper (e.g., _get_base_score(booster)) in the converter module and importing it in the test would eliminate the duplication and make any future schema changes easier to handle in one place.