Preserve dtype for NumPy array-protocol inputs

python/cuml/cuml/internals/validation.py

@@ -656,7 +656,14 @@ def check_array(
         if not isinstance(array_dtype, np.dtype) and array_dtype is not None:
             array_dtype = _as_numpy_dtype(array_dtype)
         elif not isinstance(array_dtype, np.dtype):
-            array_dtype = None
+            # Objects implementing the numpy array protocol may not expose a
+            # ``dtype`` attribute themselves. Normalize these before selecting
+            # from the supported dtypes so their represented dtype is preserved.
+            if hasattr(array, "__array__"):
+                array = np.asarray(array)
+                array_dtype = array.dtype
+            else:
+                array_dtype = None
 
     # Infer proper output dtype
     if array_dtype is not None:

python/cuml/cuml_accel_tests/upstream/scikit-learn/xfail-list.yaml

@@ -638,10 +638,7 @@
   - "sklearn.preprocessing.tests.test_polynomial::test_csr_polynomial_expansion_index_overflow[csr_array-False-True-3-2344]"
   - "sklearn.svm.tests.test_svm::test_custom_kernel_not_array_input[SVC]"
   - "sklearn.svm.tests.test_svm::test_custom_kernel_not_array_input[SVR]"
-  - "sklearn.tests.test_common::test_estimators[LinearSVC()-check_classifier_data_not_an_array]"
-  - "sklearn.tests.test_common::test_estimators[LinearSVR()-check_regressor_data_not_an_array]"
-  - "sklearn.tests.test_common::test_estimators[SVC()-check_classifier_data_not_an_array]"
-  - "sklearn.tests.test_common::test_estimators[SVR()-check_regressor_data_not_an_array]"
+  - "sklearn.tests.test_common::test_estimators[RandomForestRegressor()-check_regressor_data_not_an_array]"
   - "sklearn.tests.test_multioutput::test_base_chain_fit_and_predict_with_sparse_data_and_cv[csr_array]"
   - "sklearn.tests.test_multioutput::test_classifier_chain_fit_and_predict_with_sparse_data[csr_array]"
 - reason: Test is flaky with cuml.accel on scikit-learn <1.9
@@ -820,7 +817,6 @@
   tests:
   - "sklearn.tests.test_common::test_estimators[KNeighborsRegressor()-check_supervised_y_no_nan]"
   - "sklearn.tests.test_common::test_estimators[RandomForestClassifier()-check_classifiers_multilabel_output_format_decision_function]"
-  - "sklearn.tests.test_common::test_estimators[RandomForestRegressor()-check_regressor_data_not_an_array]"
 - reason: test_estimators checks fail
   marker: cuml_accel_test_estimators
   strict: false

python/cuml/tests/test_sklearn_compatibility.py

@@ -222,9 +222,8 @@ def _all_cuml_estimators():
     },
     RandomForestRegressor: {
         "check_regressor_data_not_an_array": (
-            "cuml defaults to float32 for non-arrays (while sklearn defaults to "
-            "float64). Our float32 and float64 results differ _just enough_ that "
-            "this test fails on tolerances."
+            "Predictions from repeated fits on equivalent inputs can differ "
+            "beyond the check's tolerance"
         ),
         "check_sample_weight_equivalence_on_dense_data": (
             "RandomForest uses quantile-binned splits, so sample weighting is "
@@ -258,9 +257,7 @@ def _all_cuml_estimators():
     },
     UMAP: {
         "check_transformer_data_not_an_array": (
-            "cuml defaults to float32 for non-arrays (while sklearn defaults to "
-            "float64). Our float32 and float64 results differ _just enough_ that "
-            "this test fails on tolerances."
+            "UMAP does not have consistent fit_transform and transform outputs"
         ),
         "check_methods_sample_order_invariance": "UMAP results depend on sample order",
         "check_transformer_general": "UMAP does not have consistent fit_transform and transform outputs",

python/cuml/tests/test_validation.py

@@ -739,6 +739,21 @@ def test_check_array_dtype(array, mem_type):
     assert out.dtype == "float32"
 
 
+@pytest.mark.parametrize("dtype", ["float32", "float64"])
+@pytest.mark.parametrize("mem_type", ["device", "host", None])
+def test_check_array_array_protocol_preserves_dtype(dtype, mem_type):
+    class ArrayLike:
+        def __init__(self, array):
+            self.array = array
+
+        def __array__(self, dtype=None, copy=None):
+            return self.array
+
+    array = ArrayLike(np.array([[1, 2, 3]], dtype=dtype))
+    out = check_array(array, dtype=("float32", "float64"), mem_type=mem_type)
+    assert out.dtype == dtype
+
+
 @example(mem_type="device", dtype="int32", order="C", shape=(3, 4))
 @example(mem_type="host", dtype="float32", order="F", shape=(3,))
 @given(