Propagate random state to numpy rng

When generating a classification problem a mix of cupy and numpy random
generators is used. The random state needs to be propagated to the numpy
generator.

Author: betatim

python/cuml/cuml/datasets/classification.py

@@ -1,4 +1,4 @@
-# Copyright (c) 2020-2024, NVIDIA CORPORATION.
+# Copyright (c) 2020-2025, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -39,18 +39,23 @@ def _generate_hypercube(samples, dimensions, rng):
                            make_classification."
         )
 
-    from sklearn.utils.random import sample_without_replacement
+    from sklearn.utils.random import (
+        sample_without_replacement,
+        check_random_state,
+    )
+
+    # Least worst way to derive a Numpy random state from a cupy random state?
+    rs = check_random_state(int(rng.randint(dimensions)))
 
     if dimensions > 30:
         return np.hstack(
             [
-                np.random.randint(2, size=(samples, dimensions - 30)),
+                rs.randint(2, size=(samples, dimensions - 30)),
                 _generate_hypercube(samples, 30, rng),
             ]
         )
-    random_state = int(rng.randint(dimensions))
     out = sample_without_replacement(
-        2**dimensions, samples, random_state=random_state
+        2**dimensions, samples, random_state=rs
     ).astype(dtype=">u4", copy=False)
     out = np.unpackbits(out.view(">u1")).reshape((-1, 32))[:, -dimensions:]
     return out

python/cuml/cuml/tests/test_make_classification.py

@@ -1,5 +1,5 @@
 #
-# Copyright (c) 2020-2024, NVIDIA CORPORATION.
+# Copyright (c) 2020-2025, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -147,3 +147,30 @@ def test_make_classification_informative_features():
         make(
             n_features=2, n_informative=2, n_classes=3, n_clusters_per_class=2
         )
+
+
+def test_make_classification_random_state():
+    # Check that results are stable across repeated calls
+
+    # We need to use more than 30 features to test all of the code paths
+    X, y = make_classification(n_features=30 + 2, random_state=42)
+    X2, y2 = make_classification(n_features=30 + 2, random_state=42)
+    assert array_equal(X, X2)
+    assert array_equal(y, y2)
+
+    # Check that results are different across different random states
+    X3, y3 = make_classification(n_features=30 + 2, random_state=43)
+    assert not array_equal(X, X3)
+    assert not array_equal(y, y3)
+
+
+def test_make_classification_random_state_gh_6510():
+    # Non regression test for gh-6510
+    X, y = make_classification(
+        10, 35, n_redundant=0, n_repeated=0, n_informative=35, random_state=42
+    )
+    X2, y2 = make_classification(
+        10, 35, n_redundant=0, n_repeated=0, n_informative=35, random_state=42
+    )
+    assert array_equal(X, X2)
+    assert array_equal(y, y2)