Merge pull request rapidsai#15800 from rapidsai/branch-24.06

Forward-merge branch-24.06 into branch-24.08

Author: GPUtester

python/cudf/cudf/core/_base_index.py

@@ -145,7 +145,7 @@ def name(self):
         raise NotImplementedError
 
     @property  # type: ignore
-    def ndim(self):  # noqa: D401
+    def ndim(self) -> int:  # noqa: D401
         """Number of dimensions of the underlying data, by definition 1."""
         return 1
 

python/cudf/cudf/core/dataframe.py

@@ -1234,7 +1234,7 @@ def dtypes(self):
         return pd.Series(self._dtypes, dtype="object")
 
     @property
-    def ndim(self):
+    def ndim(self) -> int:
         """Dimension of the data. DataFrame ndim is always 2."""
         return 2
 

python/cudf/cudf/core/frame.py

@@ -6,6 +6,7 @@
 import itertools
 import operator
 import pickle
+import types
 import warnings
 from collections import abc
 from typing import (
@@ -91,6 +92,10 @@ def _dtypes(self):
             zip(self._data.names, (col.dtype for col in self._data.columns))
         )
 
+    @property
+    def ndim(self) -> int:
+        raise NotImplementedError()
+
     @_cudf_nvtx_annotate
     def serialize(self):
         # TODO: See if self._data can be serialized outright
@@ -417,51 +422,60 @@ def __arrow_array__(self, type=None):
     @_cudf_nvtx_annotate
     def _to_array(
         self,
-        get_column_values: Callable,
-        make_empty_matrix: Callable,
+        get_array: Callable,
+        module: types.ModuleType,
+        copy: bool,
         dtype: Union[Dtype, None] = None,
         na_value=None,
-    ) -> Union[cupy.ndarray, np.ndarray]:
+    ) -> Union[cupy.ndarray, numpy.ndarray]:
         # Internal function to implement to_cupy and to_numpy, which are nearly
         # identical except for the attribute they access to generate values.
 
-        def get_column_values_na(col):
+        def to_array(
+            col: ColumnBase, dtype: np.dtype
+        ) -> Union[cupy.ndarray, numpy.ndarray]:
             if na_value is not None:
                 col = col.fillna(na_value)
-            return get_column_values(col)
+            array = get_array(col)
+            casted_array = module.asarray(array, dtype=dtype)
+            if copy and casted_array is array:
+                # Don't double copy after asarray
+                casted_array = casted_array.copy()
+            return casted_array
 
-        # Early exit for an empty Frame.
         ncol = self._num_columns
         if ncol == 0:
-            return make_empty_matrix(
-                shape=(len(self), ncol), dtype=np.dtype("float64"), order="F"
+            return module.empty(
+                shape=(len(self), ncol),
+                dtype=numpy.dtype("float64"),
+                order="F",
             )
 
         if dtype is None:
-            dtypes = [col.dtype for col in self._data.values()]
-            for dtype in dtypes:
-                if isinstance(
-                    dtype,
-                    (
-                        cudf.ListDtype,
-                        cudf.core.dtypes.DecimalDtype,
-                        cudf.StructDtype,
-                    ),
-                ):
-                    raise NotImplementedError(
-                        f"{dtype} cannot be exposed as a cupy array"
-                    )
-            dtype = find_common_type(dtypes)
+            if ncol == 1:
+                dtype = next(iter(self._data.values())).dtype
+            else:
+                dtype = find_common_type(
+                    [col.dtype for col in self._data.values()]
+                )
 
-        matrix = make_empty_matrix(
-            shape=(len(self), ncol), dtype=dtype, order="F"
-        )
-        for i, col in enumerate(self._data.values()):
-            # TODO: col.values may fail if there is nullable data or an
-            # unsupported dtype. We may want to catch and provide a more
-            # suitable error.
-            matrix[:, i] = get_column_values_na(col)
-        return matrix
+            if not isinstance(dtype, numpy.dtype):
+                raise NotImplementedError(
+                    f"{dtype} cannot be exposed as an array"
+                )
+
+        if self.ndim == 1:
+            return to_array(self._data.columns[0], dtype)
+        else:
+            matrix = module.empty(
+                shape=(len(self), ncol), dtype=dtype, order="F"
+            )
+            for i, col in enumerate(self._data.values()):
+                # TODO: col.values may fail if there is nullable data or an
+                # unsupported dtype. We may want to catch and provide a more
+                # suitable error.
+                matrix[:, i] = to_array(col, dtype)
+            return matrix
 
     # TODO: As of now, calling cupy.asarray is _much_ faster than calling
     # to_cupy. We should investigate the reasons why and whether we can provide
@@ -496,10 +510,9 @@ def to_cupy(
         cupy.ndarray
         """
         return self._to_array(
-            (lambda col: col.values.copy())
-            if copy
-            else (lambda col: col.values),
-            cupy.empty,
+            lambda col: col.values,
+            cupy,
+            copy,
             dtype,
             na_value,
         )
@@ -536,7 +549,7 @@ def to_numpy(
             )
 
         return self._to_array(
-            (lambda col: col.values_host), np.empty, dtype, na_value
+            lambda col: col.values_host, numpy, copy, dtype, na_value
         )
 
     @_cudf_nvtx_annotate

python/cudf/cudf/core/multiindex.py

@@ -563,7 +563,7 @@ def levels(self):
 
     @property  # type: ignore
     @_cudf_nvtx_annotate
-    def ndim(self):
+    def ndim(self) -> int:
         """Dimension of the data. For MultiIndex ndim is always 2."""
         return 2
 

python/cudf/cudf/core/single_column_frame.py

@@ -77,7 +77,7 @@ def name(self, value):
 
     @property  # type: ignore
     @_cudf_nvtx_annotate
-    def ndim(self):  # noqa: D401
+    def ndim(self) -> int:  # noqa: D401
         """Number of dimensions of the underlying data, by definition 1."""
         return 1