Fix cudf.pandas series accessor failures across numpy/nullable/arrow dtypes

python/cudf/cudf/core/accessors/categorical.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 from __future__ import annotations
@@ -90,11 +90,15 @@ def codes(self) -> Series:
         Return Series of codes as well as the index.
         """
         from cudf.core.series import Series
+        from cudf.utils.dtypes import min_signed_type
 
         index = (
             self._parent.index if isinstance(self._parent, Series) else None
         )
-        return Series._from_column(self._column.codes, index=index)
+        codes = self._column.codes.astype(
+            min_signed_type(len(self._column.dtype.categories))
+        )
+        return Series._from_column(codes, index=index)
 
     @property
     def ordered(self) -> bool | None:

python/cudf/cudf/core/accessors/lists.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 from __future__ import annotations
@@ -32,7 +32,8 @@ class ListMethods(BaseAccessor):
     def __init__(self, parent: Series | Index):
         if not is_dtype_obj_list(parent.dtype):
             raise AttributeError(
-                "Can only use .list accessor with a 'list' dtype"
+                "Can only use the '.list' accessor with 'list[pyarrow]' "
+                f"dtype, not {parent.dtype}."
             )
         super().__init__(parent=parent)
 

python/cudf/cudf/core/accessors/string.py

@@ -100,8 +100,23 @@ def __init__(self, parent: Series | Index):
             # Validate dtype is suitable for string operations
             is_valid_string = is_string_dtype(value_type)
         if not is_valid_string:
+            # Mirror pandas' inferred-type naming in the error message
+            # (see pandas lib.infer_dtype) so error messages match.
+            kind_to_inferred = {
+                "i": "integer",
+                "u": "integer",
+                "f": "floating",
+                "b": "boolean",
+                "c": "complex",
+                "M": "datetime64",
+                "m": "timedelta64",
+            }
+            inferred_dtype = kind_to_inferred.get(
+                getattr(value_type, "kind", ""), str(value_type)
+            )
             raise AttributeError(
-                "Can only use .str accessor with string values"
+                "Can only use .str accessor with string values, "
+                f"not {inferred_dtype}"
             )
         super().__init__(parent=parent)
 

python/cudf/cudf/core/accessors/struct.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 from __future__ import annotations
@@ -25,7 +25,8 @@ class StructMethods(BaseAccessor):
     def __init__(self, parent: Series | Index):
         if not is_dtype_obj_struct(parent.dtype):
             raise AttributeError(
-                "Can only use .struct accessor with a 'struct' dtype"
+                "Can only use the '.struct' accessor with 'struct[pyarrow]' "
+                f"dtype, not {parent.dtype}."
             )
         super().__init__(parent=parent)
 

python/cudf/cudf/core/column/datetime.py

@@ -511,7 +511,7 @@ def isocalendar(self) -> dict[str, ColumnBase]:
         # pandas' isocalendar for datetime64 returns UInt32 with NA preserved.
         is_arrow = isinstance(self.dtype, pd.ArrowDtype)
         target_dtype: DtypeObj = (
-            np.dtype(np.int64) if is_arrow else np.dtype(np.uint32)
+            np.dtype(np.int64) if is_arrow else pd.UInt32Dtype()
         )
         result = {
             field: self.strftime(
@@ -1001,6 +1001,20 @@ def tz_convert(self, tz: str | None) -> DatetimeColumn:
 
 
 class DatetimeTZColumn(DatetimeColumn):
+    def _round_dt(
+        self,
+        round_func: Callable[
+            [plc.Column, plc.datetime.RoundingFrequency], plc.Column
+        ],
+        freq: str,
+    ) -> ColumnBase:
+        # Rounding must operate on the local wall-clock time and then be
+        # re-localized to the original timezone to match pandas semantics.
+        rounded_local = cast(
+            "DatetimeColumn", self._local_time._round_dt(round_func, freq)
+        )
+        return rounded_local.tz_localize(str(self.tz))
+
     def to_pandas(
         self,
         *,

python/cudf/cudf/core/index.py

@@ -4018,8 +4018,14 @@ def microsecond(self) -> Index:
         >>> datetime_index.microsecond
         Index([0, 1, 2], dtype='int32')
         """
-        # .microsecond is already a cached_property
-        return Index._from_column(self._column.microsecond, name=self.name)
+        # libcudf's MICROSECOND component excludes the millisecond
+        # component, but pandas folds milliseconds into microsecond. Match
+        # pandas (and Series.dt.microsecond) by adding millisecond * 1000.
+        micro = self._column.microsecond
+        extra = self._column.millisecond.astype(np.dtype(np.int32)) * np.int32(
+            1000
+        )
+        return Index._from_column(micro + extra, name=self.name)
 
     @property
     @_performance_tracking

python/cudf/cudf/core/series.py

@@ -3997,8 +3997,21 @@ class BaseDatelikeProperties:
     Base accessor class for Series values.
     """
 
+    # Set via object.__setattr__ in __init__ (see __setattr__ below), so
+    # declare them here for type checkers.
+    series: Series
+    _frozen: bool
+
     def __init__(self, series: Series):
-        self.series = series
+        object.__setattr__(self, "series", series)
+        object.__setattr__(self, "_frozen", True)
+
+    def __setattr__(self, key, value):
+        # Mimic pandas' NoNewAttributesMixin: once initialized, disallow
+        # setting attributes that do not already exist.
+        if getattr(self, "_frozen", False) and not hasattr(self, key):
+            raise AttributeError(f"You cannot add any new attribute '{key}'")
+        object.__setattr__(self, key, value)
 
     def _return_result_like_self(self, column: ColumnBase) -> Series:
         """Return the method result like self.series"""
@@ -4366,10 +4379,14 @@ def dayofweek(self) -> Series:
         dtype: int16
         """
         res = self.series._column.weekday
-        # Pandas returns int64 for weekday
-        res = res.astype(
-            get_dtype_of_same_kind(self.series.dtype, np.dtype("int64"))
+        # Pandas returns int32 for numpy-backed dayofweek/day_of_week but
+        # int64 for the pyarrow-backed (ArrowDtype) variant.
+        target = (
+            np.dtype("int64")
+            if isinstance(self.series.dtype, pd.ArrowDtype)
+            else np.dtype("int32")
         )
+        res = res.astype(get_dtype_of_same_kind(self.series.dtype, target))
         return self._return_result_like_self(res)
 
     day_of_week = dayofweek
@@ -5300,7 +5317,10 @@ def seconds(self) -> Series:
         4    234000
         dtype: int64
         """
-        return self._return_result_like_self(self.series._column.seconds)
+        res = self.series._column.seconds.astype(
+            get_dtype_of_same_kind(self.series.dtype, np.dtype("int32"))
+        )
+        return self._return_result_like_self(res)
 
     @property
     @_performance_tracking
@@ -5332,7 +5352,10 @@ def microseconds(self) -> Series:
         4    234000
         dtype: int64
         """
-        return self._return_result_like_self(self.series._column.microseconds)
+        res = self.series._column.microseconds.astype(
+            get_dtype_of_same_kind(self.series.dtype, np.dtype("int32"))
+        )
+        return self._return_result_like_self(res)
 
     @property
     @_performance_tracking
@@ -5364,7 +5387,10 @@ def nanoseconds(self) -> Series:
         4    234
         dtype: int64
         """
-        return self._return_result_like_self(self.series._column.nanoseconds)
+        res = self.series._column.nanoseconds.astype(
+            get_dtype_of_same_kind(self.series.dtype, np.dtype("int32"))
+        )
+        return self._return_result_like_self(res)
 
     @property
     @_performance_tracking

python/cudf/cudf/pandas/scripts/pandas-testing-plugin.py

@@ -3867,25 +3867,11 @@ def pytest_unconfigure(config):
     "tests/scalar/timestamp/test_timestamp.py::TestTimestampProperties::test_dow_historic[0000-2-29-1]": "TODO: Add a reason for failure",
     "tests/scalar/timestamp/test_timestamp.py::TestTimestampProperties::test_dow_historic[0000-3-1-2]": "TODO: Add a reason for failure",
     "tests/scalar/timestamp/test_timestamp.py::test_negative_dates": "TODO: Add a reason for failure",
-    "tests/series/accessors/test_cat_accessor.py::TestCatAccessor::test_categorical_delegations": "TODO: Add a reason for failure",
     "tests/series/accessors/test_cat_accessor.py::TestCatAccessor::test_dt_accessor_api_for_categorical[idx0]": "TODO: Add a reason for failure",
     "tests/series/accessors/test_cat_accessor.py::TestCatAccessor::test_dt_accessor_api_for_categorical[idx1]": "AssertionError: Series are different",
-    "tests/series/accessors/test_cat_accessor.py::TestCatAccessor::test_dt_accessor_api_for_categorical[idx3]": "TODO: Add a reason for failure",
     "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_date_tz": "TODO: Add a reason for failure",
     "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_dt_accessor_limited_display_api": "TODO: Add a reason for failure",
-    "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_dt_accessor_no_new_attributes": "TODO: Add a reason for failure",
     "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_dt_accessor_updates_on_inplace": "assert 0   2018-01-01\n1   2018-01-02\n2   2018-01-01\n3   2018-01-04\n4   2018-01-05\n5   2018-01-06\n6   2018-01-07...",
-    "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_dt_namespace_accessor_datetime64[D]": "TODO: Add a reason for failure",
-    "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_dt_namespace_accessor_datetime64[ms]": "TODO: Add a reason for failure",
-    "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_dt_namespace_accessor_datetime64[s]": "TODO: Add a reason for failure",
-    "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_dt_namespace_accessor_datetime64tz": "AssertionError: Attributes of Series are different",
-    "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_dt_namespace_accessor_timedelta": "TODO: Add a reason for failure",
-    "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_dt_round_tz": "AssertionError: Series are different",
-    "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_isocalendar[input_series0-expected_output0]": "TODO: Add a reason for failure",
-    "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_isocalendar[input_series1-expected_output1]": "TODO: Add a reason for failure",
-    "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_isocalendar[input_series2-expected_output2]": "TODO: Add a reason for failure",
-    "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_isocalendar[input_series4-expected_output4]": "TODO: Add a reason for failure",
-    "tests/series/accessors/test_dt_accessor.py::TestSeriesDatetimeValues::test_isocalendar[input_series5-expected_output5]": "TODO: Add a reason for failure",
     "tests/series/accessors/test_list_accessor.py::test_list_getitem_invalid_index[list_dtype0]": "TODO: Add a reason for failure",
     "tests/series/indexing/test_datetime.py::test_indexing_over_size_cutoff_period_index": "TODO: Add a reason for failure",
     "tests/series/indexing/test_datetime.py::test_indexing_unordered": "TODO: Add a reason for failure",

python/cudf/cudf/tests/indexes/datetimeindex/test_attributes.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 import numpy as np
@@ -73,6 +73,25 @@ def test_dt_index(field):
     assert_eq(getattr(gdf_data, field), getattr(data, field), exact=False)
 
 
+def test_dti_microsecond_includes_milliseconds():
+    # DatetimeIndex.microsecond must include the whole-millisecond component
+    # like pandas: 0.001002s -> 1002 microseconds (not 2). Values >= 1000
+    # exercise the millisecond fold that plain sub-millisecond data misses.
+    data = [
+        "2010-01-01 00:00:00.001002",
+        "2010-01-01 00:00:00.123456",
+        "2010-01-01 00:00:00.000999",
+        "2010-01-01 00:00:00.000000",
+    ]
+    pidx = pd.DatetimeIndex(data)
+    gidx = cudf.from_pandas(pidx)
+
+    expect = pidx.microsecond
+    got = gidx.microsecond
+    assert_eq(expect, got)
+    assert got.dtype == np.dtype("int32")
+
+
 @pytest.mark.parametrize(
     "attr",
     [

python/cudf/cudf/tests/series/accessors/test_cat.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 from textwrap import dedent
@@ -32,6 +32,22 @@ def test_categorical_basic():
     assert str(sr) == str(pdsr)
 
 
+@pytest.mark.parametrize("num_categories", [1, 3, 200, 40000])
+def test_cat_codes_signed_dtype(num_categories):
+    # cat.codes must use a signed integer dtype matching pandas (which widens
+    # int8 -> int16 -> int32 as the category count grows), not an unsigned one.
+    categories = [f"c{i}" for i in range(num_categories)]
+    ps = pd.Series(pd.Categorical(categories, categories=categories))
+    gs = cudf.from_pandas(ps)
+
+    expect = ps.cat.codes
+    got = gs.cat.codes
+
+    assert got.dtype.kind == "i"
+    assert got.dtype == expect.dtype
+    assert_eq(expect, got)
+
+
 def test_categorical_integer():
     cat = pd.Categorical.from_codes(
         [0, -1, -1, 2, 0], categories=["a", "b", "c"]
@@ -351,8 +367,10 @@ def test_add_categories_mixed_error():
 def test_categorical_allow_nan():
     gs = cudf.Series([1, 2, np.nan, 10, np.nan, None], nan_as_null=False)
     gs = gs.astype("category")
-    expected_codes = cudf.Series([0, 1, None, 2, None, None], dtype="uint8")
+    # cat.codes uses a signed dtype to match pandas (int8 here, not uint8).
+    expected_codes = cudf.Series([0, 1, None, 2, None, None], dtype="int8")
     assert_eq(expected_codes, gs.cat.codes)
+    assert gs.cat.codes.dtype == np.dtype("int8")
 
     expected_categories = cudf.Index([1.0, 2.0, 10.0], dtype="float64")
     assert_eq(expected_categories, gs.cat.categories)

python/cudf/cudf/tests/series/accessors/test_dt.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 import datetime
 import zoneinfo
@@ -551,6 +551,65 @@ def test_isocalendar_series(data):
     assert_eq(expect, got, check_dtype=False)
 
 
+@pytest.mark.parametrize("dtype", _DT_DTYPES)
+def test_isocalendar_dtype(dtype):
+    # isocalendar's year/week/day columns must match pandas' backing dtype:
+    # UInt32 for numpy datetime, int64[pyarrow] for ArrowDtype timestamps.
+    ps = pd.Series(["2010-01-01", "2010-12-31", None], dtype=dtype)
+    gs = cudf.from_pandas(ps)
+
+    expect = ps.dt.isocalendar()
+    got = gs.dt.isocalendar()
+
+    # Exercise both value equality and the exact per-column dtype.
+    assert_eq(expect, got)
+    assert dict(got.dtypes) == dict(expect.dtypes)
+    if isinstance(dtype, pd.ArrowDtype):
+        assert all(dt == pd.ArrowDtype(pa.int64()) for dt in got.dtypes)
+    else:
+        assert all(dt == pd.UInt32Dtype() for dt in got.dtypes)
+
+
+@pytest.mark.parametrize("dtype", _DT_DTYPES)
+@pytest.mark.parametrize("attr", ["dayofweek", "day_of_week", "weekday"])
+def test_dayofweek_dtype(dtype, attr):
+    # dayofweek/day_of_week/weekday are aliases. pandas returns int32 for the
+    # numpy-backed variant but int64[pyarrow] for the ArrowDtype variant.
+    ps = pd.Series(["2010-01-01", "2010-12-31", "2011-06-15"], dtype=dtype)
+    gs = cudf.from_pandas(ps)
+
+    expect = getattr(ps.dt, attr)
+    got = getattr(gs.dt, attr)
+
+    assert_eq(expect, got)
+    if isinstance(dtype, pd.ArrowDtype):
+        assert got.dtype == pd.ArrowDtype(pa.int64())
+        assert expect.dtype == pd.ArrowDtype(pa.int64())
+    else:
+        assert got.dtype == np.dtype("int32")
+        assert expect.dtype == np.dtype("int32")
+
+
+def test_timedelta_components_dtype():
+    # pandas returns int64 for .dt.days but int32 for the sub-day components
+    # (seconds/microseconds/nanoseconds); cudf must match.
+    ps = pd.Series([1000000, 200000, 3000000], dtype="timedelta64[ns]")
+    gs = cudf.from_pandas(ps)
+
+    for attr, expected_dtype in [
+        ("days", np.dtype("int64")),
+        ("seconds", np.dtype("int32")),
+        ("microseconds", np.dtype("int32")),
+        ("nanoseconds", np.dtype("int32")),
+    ]:
+        expect = getattr(ps.dt, attr)
+        got = getattr(gs.dt, attr)
+        assert_eq(expect, got)
+        assert got.dtype == expected_dtype
+        # Sanity: our expectation tracks pandas' actual dtype.
+        assert expect.dtype == expected_dtype
+
+
 @pytest.mark.parametrize(
     "data", [[1, 2, 3, None], [], [100121, 1221312, 321312321, 1232131223]]
 )