Merge pull request #17821 from rapidsai/branch-25.02

Forward-merge branch-25.02 into branch-25.04

python/cudf/cudf/core/column/struct.py

@@ -267,13 +267,15 @@ def explode(self):
         2  3  z
         3  4  a
         """
+        data = {
+            name: col.copy(deep=True)
+            for name, col in zip(
+                self._column.dtype.fields, self._column.children
+            )
+        }
+        rangeindex = len(data) == 0
         return cudf.DataFrame._from_data(
             cudf.core.column_accessor.ColumnAccessor(
-                {
-                    name: col.copy(deep=True)
-                    for name, col in zip(
-                        self._column.dtype.fields, self._column.children
-                    )
-                }
+                data, rangeindex=rangeindex
             )
         )

python/cudf/cudf/core/dataframe.py

@@ -789,6 +789,8 @@ def __init__(
                     label_dtype=label_dtype,
                     verify=False,
                 )
+            else:
+                self._data.rangeindex = True
         elif isinstance(data, ColumnAccessor):
             raise TypeError(
                 "Use cudf.DataFrame._from_data for constructing a DataFrame from "
@@ -1120,6 +1122,9 @@ def _init_from_dict_like(
                     data[col_name],
                     nan_as_null=nan_as_null,
                 )
+        elif columns is None:
+            self._data.rangeindex = True
+
         self._data._level_names = (
             tuple(columns.names)
             if isinstance(columns, pd.Index)

python/cudf/cudf/io/avro.py

@@ -1,9 +1,10 @@
-# Copyright (c) 2019-2024, NVIDIA CORPORATION.
+# Copyright (c) 2019-2025, NVIDIA CORPORATION.
 
 import pylibcudf as plc
 
 import cudf
 from cudf._lib.column import Column
+from cudf.core.column_accessor import ColumnAccessor
 from cudf.utils import ioutils
 
 
@@ -54,4 +55,5 @@ def read_avro(
             strict=True,
         )
     }
-    return cudf.DataFrame._from_data(data)
+    ca = ColumnAccessor(data, rangeindex=len(data) == 0)
+    return cudf.DataFrame._from_data(ca)

python/cudf/cudf/io/csv.py

@@ -18,6 +18,7 @@
 from cudf._lib.column import Column
 from cudf.api.types import is_hashable, is_scalar
 from cudf.core.buffer import acquire_spill_lock
+from cudf.core.column_accessor import ColumnAccessor
 from cudf.utils import ioutils
 from cudf.utils.dtypes import (
     _maybe_convert_to_default_type,
@@ -262,8 +263,8 @@ def read_csv(
             strict=True,
         )
     }
-
-    df = cudf.DataFrame._from_data(data)
+    ca = ColumnAccessor(data, rangeindex=len(data) == 0)
+    df = cudf.DataFrame._from_data(ca)
 
     if isinstance(dtype, abc.Mapping):
         for k, v in dtype.items():

python/cudf/cudf/io/orc.py

@@ -13,6 +13,7 @@
 from cudf._lib.column import Column
 from cudf.api.types import is_list_like
 from cudf.core.buffer import acquire_spill_lock
+from cudf.core.column_accessor import ColumnAccessor
 from cudf.core.index import _index_from_data
 from cudf.utils import ioutils
 from cudf.utils.dtypes import dtype_to_pylibcudf_type
@@ -267,7 +268,7 @@ def read_orc(
             # When `columns=[]`, index needs to be
             # established, but not the columns.
             nrows = tbl_w_meta.tbl.num_rows()
-            data = {}
+            ca = ColumnAccessor({})
             index = cudf.RangeIndex(nrows)
         else:
             names = tbl_w_meta.column_names(include_children=False)
@@ -367,8 +368,9 @@ def read_orc(
                     data.items(), child_name_values
                 )
             }
+            ca = ColumnAccessor(data, rangeindex=len(data) == 0)
 
-        return cudf.DataFrame._from_data(data, index=index)
+        return cudf.DataFrame._from_data(ca, index=index)
     else:
         from pyarrow import orc
 

python/cudf/cudf/tests/dataframe/test_io_serialization.py

@@ -1,5 +1,4 @@
-# Copyright (c) 2023-2024, NVIDIA CORPORATION.
-import contextlib
+# Copyright (c) 2023-2025, NVIDIA CORPORATION.
 from io import BytesIO
 
 import pandas as pd
@@ -35,18 +34,11 @@ def test_dataframe_parquet_roundtrip(index, write_index, empty):
     metadata_equal = (
         gpu_table.schema.pandas_metadata == cpu_table.schema.pandas_metadata
     )
-    if empty and write_index is not False:
-        # https://github.com/rapidsai/cudf/issues/15372
-        ctx = pytest.raises(AssertionError)
-    else:
-        ctx = contextlib.nullcontext()
-    with ctx:
-        assert metadata_equal
+    assert metadata_equal
 
     gpu_read = cudf.read_parquet(gpu_buf)
     cpu_read = cudf.read_parquet(cpu_buf)
-    with ctx:
-        assert_eq(gpu_read, cpu_read)
+    assert_eq(gpu_read, cpu_read)
 
 
 @pytest.mark.parametrize("preserve_index", [False, True, None])

python/cudf/cudf/tests/test_dataframe.py

@@ -8811,7 +8811,9 @@ def test_dataframe_mode(df, numeric_only, dropna):
 
     expected = pdf.mode(numeric_only=numeric_only, dropna=dropna)
     actual = df.mode(numeric_only=numeric_only, dropna=dropna)
-
+    if len(actual.columns) == 0:
+        # pandas < 3.0 returns an Index[object] instead of RangeIndex
+        actual.columns = expected.columns
     assert_eq(expected, actual, check_dtype=False)
 
 
@@ -11260,3 +11262,10 @@ def test_roundtrip_dataframe_plc_table(pdf):
     expect = cudf.DataFrame.from_pandas(pdf)
     actual = cudf.DataFrame.from_pylibcudf(*expect.to_pylibcudf())
     assert_eq(expect, actual)
+
+
+@pytest.mark.parametrize("data", [None, {}])
+def test_empty_construction_rangeindex_columns(data):
+    result = cudf.DataFrame(data=data).columns
+    expected = pd.RangeIndex(0)
+    pd.testing.assert_index_equal(result, expected, exact=True)