Check for invalid cuDF input. (#11248)

python-package/xgboost/_data_utils.py

@@ -16,6 +16,11 @@ class _ArrayLikeArg(Protocol):
     def __array_interface__(self) -> "ArrayInf": ...
 
 
+class _CudaArrayLikeArg(Protocol):
+    @property
+    def __cuda_array_interface__(self) -> "ArrayInf": ...
+
+
 class TransformedDf(Protocol):
     """Protocol class for storing transformed dataframe."""
 
@@ -151,3 +156,12 @@ def array_interface(data: np.ndarray) -> bytes:
     interface = array_interface_dict(data)
     interface_str = bytes(json.dumps(interface), "utf-8")
     return interface_str
+
+
+def check_cudf_meta(data: _CudaArrayLikeArg, field: str) -> None:
+    "Make sure no missing value in meta data."
+    if (
+        "mask" in data.__cuda_array_interface__
+        and data.__cuda_array_interface__["mask"] is not None
+    ):
+        raise ValueError(f"Missing value is not allowed for: {field}")

python-package/xgboost/data.py

@@ -29,6 +29,7 @@
     array_hasobject,
     array_interface,
     array_interface_dict,
+    check_cudf_meta,
     cuda_array_interface,
     make_array_interface,
 )
@@ -1555,14 +1556,15 @@ def _meta_from_cudf_df(data: DataType, field: str, handle: ctypes.c_void_p) -> N
 
 
 def _meta_from_cudf_series(data: DataType, field: str, handle: ctypes.c_void_p) -> None:
-    interface = bytes(json.dumps([data.__cuda_array_interface__], indent=2), "utf-8")
-    _check_call(_LIB.XGDMatrixSetInfoFromInterface(handle, c_str(field), interface))
+    check_cudf_meta(data, field)
+    inf = cuda_array_interface(data)
+    _check_call(_LIB.XGDMatrixSetInfoFromInterface(handle, c_str(field), inf))
 
 
 def _meta_from_cupy_array(data: DataType, field: str, handle: ctypes.c_void_p) -> None:
     data = _transform_cupy_array(data)
-    interface = bytes(json.dumps([data.__cuda_array_interface__], indent=2), "utf-8")
-    _check_call(_LIB.XGDMatrixSetInfoFromInterface(handle, c_str(field), interface))
+    inf = cuda_array_interface(data)
+    _check_call(_LIB.XGDMatrixSetInfoFromInterface(handle, c_str(field), inf))
 
 
 def dispatch_meta_backend(
@@ -1604,15 +1606,15 @@ def dispatch_meta_backend(
         data = _transform_dlpack(data)
         _meta_from_cupy_array(data, name, handle)
         return
-    if _is_cupy_alike(data):
-        _meta_from_cupy_array(data, name, handle)
-        return
     if _is_cudf_ser(data):
         _meta_from_cudf_series(data, name, handle)
         return
     if _is_cudf_df(data):
         _meta_from_cudf_df(data, name, handle)
         return
+    if _is_cupy_alike(data):
+        _meta_from_cupy_array(data, name, handle)
+        return
     if _is_modin_df(data):
         _meta_from_pandas_df(data, name, dtype=dtype, handle=handle)
         return

tests/python-gpu/test_from_cudf.py

@@ -382,3 +382,20 @@ def test_from_cudf_iter(enable_categorical):
     predict = reg.predict(m)
     predict_with_it = reg_with_it.predict(m_it)
     np.testing.assert_allclose(predict_with_it, predict)
+
+
+def test_invalid_meta() -> None:
+    df = cudf.DataFrame({"f0": [0, 1, 2], "f1": [2, 3, 4], "y": [None, 1, 2]})
+    y = df["y"]
+    X = df.drop(["y"], axis=1)
+    with pytest.raises(ValueError, match="Missing value"):
+        xgb.DMatrix(X, y)
+    with pytest.raises(ValueError, match="Missing value"):
+        xgb.QuantileDMatrix(X, y)
+    y = X.copy()
+    y.iloc[0, 0] = None
+    # check by the cuDF->cupy converter.
+    with pytest.raises(ValueError, match="no nulls"):
+        xgb.DMatrix(X, y)
+    with pytest.raises(ValueError, match="no nulls"):
+        xgb.QuantileDMatrix(X, y)