Add dedicated handling for cudf.pandas wrapped Numpy arrays

python/cuml/internals/input_utils.py

@@ -174,6 +174,12 @@ def get_supported_input_type(X):
     if isinstance(X, CudfIndex):
         return CudfIndex
 
+    # A cudf.pandas wrapped Numpy array defines `__cuda_array_interface__`
+    # which means without this we'd always return a cupy array. We don't want
+    # to match wrapped cupy arrays, they get dealt with later
+    if getattr(X, "_fsproxy_slow_type", None) is np.ndarray:
+        return np.ndarray
+
     try:
         if numba_cuda.devicearray.is_cuda_ndarray(X):
             return numba_cuda.devicearray.DeviceNDArrayBase

python/cuml/tests/conftest.py

@@ -1,5 +1,5 @@
 #
-# Copyright (c) 2018-2023, NVIDIA CORPORATION.
+# Copyright (c) 2018-2024, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -30,6 +30,7 @@
 import os
 import subprocess
 import pandas as pd
+import cudf.pandas
 
 from cuml.internals.safe_imports import cpu_only_import
 
@@ -169,6 +170,10 @@ def pytest_collection_modifyitems(config, items):
 
 
 def pytest_configure(config):
+    config.addinivalue_line(
+        "markers",
+        "cudf_pandas: mark test as requiring the cudf.pandas wrapper",
+    )
     cp.cuda.set_allocator(None)
     # max_gpu_memory: Capacity of the GPU memory in GB
     pytest.max_gpu_memory = get_gpu_memory()
@@ -186,6 +191,16 @@ def pytest_configure(config):
         hypothesis.settings.load_profile("unit")
 
 
+def pytest_pyfunc_call(pyfuncitem):
+    """Skip tests that require the cudf.pandas accelerator
+
+    Tests marked with `@pytest.mark.cudf_pandas` will only be run if the
+    cudf.pandas accelerator is enabled via the `cudf.pandas` plugin.
+    """
+    if "cudf_pandas" in pyfuncitem.keywords and not cudf.pandas.LOADED:
+        pytest.skip("Test requires cudf.pandas accelerator")
+
+
 @pytest.fixture(scope="module")
 def nlp_20news():
     try:

python/cuml/tests/test_input_utils.py

@@ -1,5 +1,5 @@
 #
-# Copyright (c) 2019-2023, NVIDIA CORPORATION.
+# Copyright (c) 2019-2024, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -14,7 +14,9 @@
 # limitations under the License.
 #
 
+import numpy as np
 from pandas import Series as pdSeries
+from cuml.manifold import umap
 from cuml.internals.safe_imports import cpu_only_import_from
 from cuml.internals.safe_imports import gpu_only_import_from
 from cuml.internals.input_utils import convert_dtype
@@ -24,6 +26,7 @@
 from cuml.common import input_to_cuml_array, CumlArray
 from cuml.internals.safe_imports import cpu_only_import
 import pytest
+import pandas as pd
 
 from cuml.internals.safe_imports import gpu_only_import
 
@@ -442,3 +445,18 @@ def test_tocupy_missing_values_handling():
         array, n_rows, n_cols, dtype = input_to_cupy_array(
             df, fail_on_null=True
         )
+
+
+@pytest.mark.cudf_pandas
+def test_numpy_output():
+    # Check that a Numpy array is used as output when a cudf.pandas wrapped
+    # Numpy array is passed in.
+    # Non regression test for issue #5784
+    df = pd.DataFrame({"a": range(5), "b": range(5)})
+    X = df.values
+
+    reducer = umap.UMAP()
+
+    # Check that this is a cudf.pandas wrapped array
+    assert hasattr(X, "_fsproxy_fast_type")
+    assert isinstance(reducer.fit_transform(X), np.ndarray)