Incompatibility Between QuantileDMatrix and xgboost.cv Function

[hugocool]

I have encountered an issue when attempting to perform cross-validation using the xgboost.cv function with a QuantileDMatrix as input. The operation results in the following error:

XGBoostError: [17:29:55] /workspace/src/data/iterative_dmatrix.h:88: Slicing DMatrix is not supported for Quantile DMatrix.
Stack trace:
  [bt] (0) /opt/pysetup/.venv/lib/python3.10/site-packages/xgboost/lib/libxgboost.so(+0x22dcbc) [0x7f7553e2dcbc]
  [bt] (1) /opt/pysetup/.venv/lib/python3.10/site-packages/xgboost/lib/libxgboost.so(+0x525d31) [0x7f7554125d31]
  [bt] (2) /opt/pysetup/.venv/lib/python3.10/site-packages/xgboost/lib/libxgboost.so(XGDMatrixSliceDMatrixEx+0x15b) [0x7f7553d448cb]
  [bt] (3) /usr/lib/x86_64-linux-gnu/libffi.so.8(+0x7e2e) [0x7f76ce294e2e]
  [bt] (4) /usr/lib/x86_64-linux-gnu/libffi.so.8(+0x4493) [0x7f76ce291493]
  [bt] (5) /opt/pyenv/versions/3.10.12/lib/python3.10/lib-dynload/_ctypes.cpython-310-x86_64-linux-gnu.so(+0x12ea5) [0x7f76cdcf0ea5]
  [bt] (6) /opt/pyenv/versions/3.10.12/lib/python3.10/lib-dynload/_ctypes.cpython-310-x86_64-linux-gnu.so(+0xc940) [0x7f76cdcea940]
  [bt] (7) /opt/pyenv/versions/3.10.12/lib/libpython3.10.so.1.0(_PyObject_MakeTpCall+0x8c) [0x7f76d4175e8c]
  [bt] (8) /opt/pyenv/versions/3.10.12/lib/libpython3.10.so.1.0(_PyEval_EvalFrameDefault+0x8f84) [0x7f76d4122ed4]

Steps to Reproduce:

1. Create a QuantileDMatrix from a dataset.
2. Attempt to perform cross-validation using xgboost.cv with this QuantileDMatrix.

Expected Behavior:

The xgboost.cv function should either:

• Support cross-validation with QuantileDMatrix.
• Clearly document that QuantileDMatrix is incompatible with xgboost.cv and provide guidance on alternative approaches.

Environment Details:

• XGBoost version: 2.1.3
• Python version: 3.10
• Operating System: ubuntu 24

Additional Information:

This issue arises because QuantileDMatrix does not support slicing operations, which are integral to the cross-validation process in xgboost.cv. While alternative approaches exist, such as using the standard DMatrix or implementing custom cross-validation, it would be beneficial for users if the library either supported this functionality or provided clear documentation regarding this limitation.

[trivialfis]

Thank you for raising the issue. We are looking for suggestions on how to design a new CV interface #11000