Check __cuda_array_interface__ instead of cupy class. (#9971)

* Now XGBoost can directly consume CUDA data from torch.

doc/python/python_intro.rst

@@ -162,6 +162,8 @@ Support Matrix
 +-------------------------+-----------+-------------------+-----------+-----------+--------------------+-------------+
 | cupy.ndarray            | T         | T                 | T         | T         | T                  | T           |
 +-------------------------+-----------+-------------------+-----------+-----------+--------------------+-------------+
+| torch.Tensor            | T         | T                 | T         | T         | T                  | T           |
++-------------------------+-----------+-------------------+-----------+-----------+--------------------+-------------+
 | dlpack                  | CPA       | CPA               |           | CPA       | FF                 | FF          |
 +-------------------------+-----------+-------------------+-----------+-----------+--------------------+-------------+
 | datatable.Frame         | T         | FF                |           | NPA       | FF                 |             |

python-package/xgboost/compat.py

@@ -138,9 +138,9 @@ def concat(value: Sequence[_T]) -> _T:  # pylint: disable=too-many-return-statem
         from cudf import concat as CUDF_concat  # pylint: disable=import-error
 
         return CUDF_concat(value, axis=0)
-    from .data import _is_cupy_array
+    from .data import _is_cupy_alike
 
-    if _is_cupy_array(value[0]):
+    if _is_cupy_alike(value[0]):
         import cupy  # pylint: disable=import-error
 
         # pylint: disable=c-extension-no-member,no-member

python-package/xgboost/core.py

@@ -357,10 +357,13 @@ def _numpy2ctypes_type(dtype: Type[np.number]) -> Type[CNumeric]:
     return _NUMPY_TO_CTYPES_MAPPING[dtype]
 
 
+def _array_hasobject(data: DataType) -> bool:
+    return hasattr(data.dtype, "hasobject") and data.dtype.hasobject
+
+
 def _cuda_array_interface(data: DataType) -> bytes:
-    assert (
-        data.dtype.hasobject is False
-    ), "Input data contains `object` dtype.  Expecting numeric data."
+    if _array_hasobject(data):
+        raise ValueError("Input data contains `object` dtype.  Expecting numeric data.")
     interface = data.__cuda_array_interface__
     if "mask" in interface:
         interface["mask"] = interface["mask"].__cuda_array_interface__
@@ -2102,7 +2105,7 @@ def boost(
             _array_interface,
             _cuda_array_interface,
             _ensure_np_dtype,
-            _is_cupy_array,
+            _is_cupy_alike,
         )
 
         self._assign_dmatrix_features(dtrain)
@@ -2116,7 +2119,7 @@ def array_interface(array: NumpyOrCupy) -> bytes:
                 "Expecting `np.ndarray` or `cupy.ndarray` for gradient and hessian."
                 f" Got: {type(array)}"
             )
-            if not isinstance(array, np.ndarray) and not _is_cupy_array(array):
+            if not isinstance(array, np.ndarray) and not _is_cupy_alike(array):
                 raise TypeError(msg)
 
             n_samples = dtrain.num_row()
@@ -2131,7 +2134,7 @@ def array_interface(array: NumpyOrCupy) -> bytes:
             if isinstance(array, np.ndarray):
                 array, _ = _ensure_np_dtype(array, array.dtype)
                 interface = _array_interface(array)
-            elif _is_cupy_array(array):
+            elif _is_cupy_alike(array):
                 interface = _cuda_array_interface(array)
             else:
                 raise TypeError(msg)
@@ -2461,7 +2464,7 @@ def inplace_predict(
             _arrow_transform,
             _is_arrow,
             _is_cudf_df,
-            _is_cupy_array,
+            _is_cupy_alike,
             _is_list,
             _is_np_array_like,
             _is_pandas_df,
@@ -2543,7 +2546,7 @@ def inplace_predict(
                 )
             )
             return _prediction_output(shape, dims, preds, False)
-        if _is_cupy_array(data):
+        if _is_cupy_alike(data):
             from .data import _transform_cupy_array
 
             data = _transform_cupy_array(data)

python-package/xgboost/dask/__init__.py

@@ -75,7 +75,7 @@
     _deprecate_positional_args,
     _expect,
 )
-from xgboost.data import _is_cudf_ser, _is_cupy_array
+from xgboost.data import _is_cudf_ser, _is_cupy_alike
 from xgboost.sklearn import (
     XGBClassifier,
     XGBClassifierBase,
@@ -1909,7 +1909,7 @@ async def _fit_async(
             self.classes_ = await self.client.compute(y.drop_duplicates())
         if _is_cudf_ser(self.classes_):
             self.classes_ = self.classes_.to_cupy()
-        if _is_cupy_array(self.classes_):
+        if _is_cupy_alike(self.classes_):
             self.classes_ = self.classes_.get()
         self.classes_ = numpy.array(self.classes_)
         self.n_classes_ = len(self.classes_)

python-package/xgboost/data.py

@@ -26,6 +26,7 @@
     DataIter,
     DataSplitMode,
     DMatrix,
+    _array_hasobject,
     _check_call,
     _cuda_array_interface,
     _ProxyDMatrix,
@@ -77,9 +78,8 @@ def is_scipy_csr(data: DataType) -> bool:
 
 
 def _array_interface_dict(data: np.ndarray) -> dict:
-    assert (
-        data.dtype.hasobject is False
-    ), "Input data contains `object` dtype.  Expecting numeric data."
+    if _array_hasobject(data):
+        raise ValueError("Input data contains `object` dtype.  Expecting numeric data.")
     interface = data.__array_interface__
     if "mask" in interface:
         interface["mask"] = interface["mask"].__array_interface__
@@ -219,7 +219,7 @@ def _is_np_array_like(data: DataType) -> bool:
 def _ensure_np_dtype(
     data: DataType, dtype: Optional[NumpyDType]
 ) -> Tuple[np.ndarray, Optional[NumpyDType]]:
-    if data.dtype.hasobject or data.dtype in [np.float16, np.bool_]:
+    if _array_hasobject(data) or data.dtype in [np.float16, np.bool_]:
         dtype = np.float32
         data = data.astype(dtype, copy=False)
     if not data.flags.aligned:
@@ -998,19 +998,16 @@ def _is_cudf_ser(data: DataType) -> bool:
     return lazy_isinstance(data, "cudf.core.series", "Series")
 
 
-def _is_cupy_array(data: DataType) -> bool:
-    return any(
-        lazy_isinstance(data, n, "ndarray")
-        for n in ("cupy.core.core", "cupy", "cupy._core.core")
-    )
+def _is_cupy_alike(data: DataType) -> bool:
+    return hasattr(data, "__cuda_array_interface__")
 
 
 def _transform_cupy_array(data: DataType) -> CupyT:
     import cupy  # pylint: disable=import-error
 
     if not hasattr(data, "__cuda_array_interface__") and hasattr(data, "__array__"):
         data = cupy.array(data, copy=False)
-    if data.dtype.hasobject or data.dtype in [cupy.bool_]:
+    if _array_hasobject(data) or data.dtype in [cupy.bool_]:
         data = data.astype(cupy.float32, copy=False)
     return data
 
@@ -1222,7 +1219,7 @@ def dispatch_data_backend(
         return _from_cudf_df(
             data, missing, threads, feature_names, feature_types, enable_categorical
         )
-    if _is_cupy_array(data):
+    if _is_cupy_alike(data):
         return _from_cupy_array(data, missing, threads, feature_names, feature_types)
     if _is_cupy_csr(data):
         raise TypeError("cupyx CSR is not supported yet.")
@@ -1354,7 +1351,7 @@ def dispatch_meta_backend(
         data = _transform_dlpack(data)
         _meta_from_cupy_array(data, name, handle)
         return
-    if _is_cupy_array(data):
+    if _is_cupy_alike(data):
         _meta_from_cupy_array(data, name, handle)
         return
     if _is_cudf_ser(data):
@@ -1419,7 +1416,7 @@ def _proxy_transform(
         return _transform_cudf_df(
             data, feature_names, feature_types, enable_categorical
         )
-    if _is_cupy_array(data):
+    if _is_cupy_alike(data):
         data = _transform_cupy_array(data)
         return data, None, feature_names, feature_types
     if _is_dlpack(data):
@@ -1470,7 +1467,7 @@ def dispatch_proxy_set_data(
         # pylint: disable=W0212
         proxy._set_data_from_cuda_columnar(data, cast(List, cat_codes))
         return
-    if _is_cupy_array(data):
+    if _is_cupy_alike(data):
         proxy._set_data_from_cuda_interface(data)  # pylint: disable=W0212
         return
     if _is_dlpack(data):

python-package/xgboost/sklearn.py

@@ -39,7 +39,7 @@
     _deprecate_positional_args,
     _parse_eval_str,
 )
-from .data import _is_cudf_df, _is_cudf_ser, _is_cupy_array, _is_pandas_df
+from .data import _is_cudf_df, _is_cudf_ser, _is_cupy_alike, _is_pandas_df
 from .training import train
 
 
@@ -1177,7 +1177,7 @@ def predict(
                         base_margin=base_margin,
                         validate_features=validate_features,
                     )
-                    if _is_cupy_array(predts):
+                    if _is_cupy_alike(predts):
                         import cupy  # pylint: disable=import-error
 
                         predts = cupy.asnumpy(predts)  # ensure numpy array is used.
@@ -1458,7 +1458,7 @@ def fit(
                 classes = cp.unique(y.values)
                 self.n_classes_ = len(classes)
                 expected_classes = cp.array(self.classes_)
-            elif _is_cupy_array(y):
+            elif _is_cupy_alike(y):
                 import cupy as cp  # pylint: disable=E0401
 
                 classes = cp.unique(y)