Deprecate max_depth=16 in RandomForest and add None support

python/cuml/cuml/dask/ensemble/base.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2021-2025, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2021-2026, NVIDIA CORPORATION.
 # SPDX-License-Identifier: Apache-2.0
 #
 
@@ -46,6 +46,17 @@ def _create_model(
             )
         self.workers = workers
         self._set_internal_model(None)
+        # TODO(26.08): Drop along with the single-GPU deprecation in
+        # cuml.ensemble.randomforest_common.
+        # Record whether the user explicitly set `max_depth`; the warning is
+        # emitted from `_fit` so it fires at fit time (matching the single-GPU
+        # path) rather than at construction. We still forward an explicit
+        # `max_depth=16` to the per-worker single-GPU estimators so they don't
+        # each emit their own FutureWarning.
+        self._max_depth_user_set = "max_depth" in kwargs
+        if not self._max_depth_user_set:
+            kwargs["max_depth"] = 16
+
         self.active_workers = list()
         self.ignore_empty_partitions = ignore_empty_partitions
         self.n_estimators = n_estimators
@@ -89,6 +100,17 @@ def _estimators_per_worker(self, n_estimators):
         return n_estimators_per_worker
 
     def _fit(self, model, dataset, convert_dtype, broadcast_data):
+        # TODO(26.08): Drop along with the single-GPU deprecation in
+        # cuml.ensemble.randomforest_common.
+        if not getattr(self, "_max_depth_user_set", True):
+            warnings.warn(
+                "The default value of 'max_depth' will change from 16 to "
+                "None (unlimited depth) in release 26.08. To suppress this "
+                "warning, set 'max_depth' explicitly.",
+                FutureWarning,
+                stacklevel=3,
+            )
+
         data = DistributedDataHandler.create(dataset, client=self.client)
         self.active_workers = data.workers
         self.datatype = data.datatype
@@ -230,6 +252,10 @@ def _get_params(self, deep):
         return params_of_each_model
 
     def _set_params(self, **params):
+        # TODO(26.08): Drop along with the single-GPU deprecation in
+        # cuml.ensemble.randomforest_common.
+        if "max_depth" in params:
+            self._max_depth_user_set = True
         model_params = list()
         for idx, worker in enumerate(self.workers):
             model_params.append(

python/cuml/cuml/dask/ensemble/randomforestclassifier.py

@@ -74,6 +74,9 @@ class RandomForestClassifier(
 
         .. note:: This default differs from scikit-learn's
           random forest, which defaults to unlimited depth.
+
+        .. versionchanged:: 26.08
+          The default of `max_depth` will change from `16` to `None`.
     max_leaves : int (default = -1)
         Maximum leaf nodes per tree. Soft constraint. Unlimited, If ``-1``.
     max_features : float (default = 'auto')

python/cuml/cuml/dask/ensemble/randomforestregressor.py

@@ -64,6 +64,9 @@ class RandomForestRegressor(
 
         .. note:: This default differs from scikit-learn's
           random forest, which defaults to unlimited depth.
+
+        .. versionchanged:: 26.08
+          The default of `max_depth` will change from `16` to `None`.
     max_leaves : int (default = -1)
         Maximum leaf nodes per tree. Soft constraint. Unlimited, If ``-1``.
     max_features : float (default = 'auto')

python/cuml/cuml/ensemble/randomforest_common.pyx

@@ -167,6 +167,9 @@ def compute_max_features(
         )
 
 
+_DEPRECATED_MAX_DEPTH_DEFAULT = "deprecated"
+
+
 class BaseRandomForestModel(Base, InteropMixin):
 
     @classmethod
@@ -244,7 +247,9 @@ class BaseRandomForestModel(Base, InteropMixin):
         return {
             "n_estimators": self.n_estimators,
             "criterion": criterion,
-            "max_depth": self.max_depth,
+            "max_depth": (
+                16 if self.max_depth == _DEPRECATED_MAX_DEPTH_DEFAULT else self.max_depth
+            ),
             "min_samples_split": self.min_samples_split,
             "min_samples_leaf": self.min_samples_leaf,
             "max_features": self.max_features,
@@ -307,7 +312,7 @@ class BaseRandomForestModel(Base, InteropMixin):
         n_estimators=100,
         bootstrap=True,
         max_samples=1.0,
-        max_depth=16,
+        max_depth=_DEPRECATED_MAX_DEPTH_DEFAULT,
         max_leaves=-1,
         max_features='sqrt',
         n_bins=128,
@@ -417,15 +422,25 @@ class BaseRandomForestModel(Base, InteropMixin):
         cdef int n_classes = self.n_classes_ if is_classifier else 0
 
         cdef int max_depth_c
-        if self.max_depth is None:
+        max_depth = self.max_depth
+
+        if max_depth == _DEPRECATED_MAX_DEPTH_DEFAULT:
+            warnings.warn(
+                "The default value of 'max_depth' will change from 16 to "
+                "None (unlimited depth) in release 26.08. To suppress this "
+                "warning, set 'max_depth' explicitly.",
+                FutureWarning, stacklevel=3)
+            max_depth = 16
+
+        if max_depth is None:
             max_depth_c = np.iinfo(np.int32).max
-        elif not isinstance(self.max_depth, int) or self.max_depth <= 0:
+        elif not isinstance(max_depth, int) or max_depth <= 0:
             raise ValueError(
                 f"max_depth must be a positive integer or None (unlimited); "
-                f"got {self.max_depth!r}"
+                f"got {max_depth!r}"
             )
         else:
-            max_depth_c = self.max_depth
+            max_depth_c = max_depth
 
         # Validate OOB score parameter
         if callable(self.oob_score):

python/cuml/cuml/ensemble/randomforestclassifier.py

@@ -76,6 +76,9 @@ class RandomForestClassifier(BaseRandomForestModel, ClassifierMixin):
 
         .. note:: This default differs from scikit-learn's random forest,
           which defaults to unlimited depth.
+
+        .. versionchanged:: 26.08
+          The default of `max_depth` will change from `16` to `None`.
     max_leaves : int (default = -1)
         Maximum leaf nodes per tree. Soft constraint. Unlimited,
         If ``-1``.

python/cuml/cuml/ensemble/randomforestregressor.py

@@ -71,6 +71,9 @@ class RandomForestRegressor(BaseRandomForestModel, RegressorMixin):
 
         .. note:: This default differs from scikit-learn's random forest,
           which defaults to unlimited depth.
+
+        .. versionchanged:: 26.08
+          The default of `max_depth` will change from `16` to `None`.
     max_leaves : int (default = -1)
         Maximum leaf nodes per tree. Soft constraint. Unlimited,
         If ``-1``.

python/cuml/tests/dask/test_dask_random_forest.py

@@ -24,6 +24,11 @@
 from cuml.ensemble import RandomForestClassifier as cuRFC_sg
 from cuml.ensemble import RandomForestRegressor as cuRFR_sg
 
+# TODO(26.08): Remove this filter
+pytestmark = pytest.mark.filterwarnings(
+    "ignore:The default value of 'max_depth':FutureWarning"
+)
+
 
 def _prep_training_data(c, X_train, y_train, partitions_per_worker):
     workers = c.has_what().keys()
@@ -234,7 +239,11 @@ def test_rf_classification_dask_fil_predict_proba(
     y_proba[:, 1] = y_test
     y_proba[:, 0] = 1.0 - y_test
     fil_mse = mean_squared_error(y_proba, fil_preds_proba)
-    sk_model = skrfc(n_estimators=40, max_depth=16, random_state=10)
+    sk_model = skrfc(
+        n_estimators=cu_rf_params["n_estimators"],
+        max_depth=cu_rf_params["max_depth"],
+        random_state=10,
+    )
     sk_model.fit(X_train, y_train)
     sk_preds_proba = sk_model.predict_proba(X_test)
     sk_mse = mean_squared_error(y_proba, sk_preds_proba)
@@ -258,7 +267,7 @@ def test_rf_concatenation_dask(client, model_type):
     else:
         y = y.astype(np.float32)
     n_estimators = 40
-    cu_rf_params = {"n_estimators": n_estimators}
+    cu_rf_params = {"n_estimators": n_estimators, "max_depth": 16}
 
     X_df, y_df = _prep_training_data(client, X, y, partitions_per_worker=2)
 
@@ -284,7 +293,8 @@ def test_single_input_regression(client, ignore_empty_partitions):
 
     X, y = _prep_training_data(client, X, y, partitions_per_worker=2)
     cu_rf_mg = cuRFR_mg(
-        n_bins=1, ignore_empty_partitions=ignore_empty_partitions
+        n_bins=1,
+        ignore_empty_partitions=ignore_empty_partitions,
     )
 
     if (

python/cuml/tests/explainer/test_explainer_common.py

@@ -1,5 +1,5 @@
 #
-# SPDX-FileCopyrightText: Copyright (c) 2020-2025, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2020-2026, NVIDIA CORPORATION.
 # SPDX-License-Identifier: Apache-2.0
 #
 
@@ -21,6 +21,10 @@
 )
 from cuml.testing.utils import ClassEnumerator
 
+# TODO(26.08) Remove this filter
+pytestmark = pytest.mark.filterwarnings(
+    "ignore:The default value of 'max_depth':FutureWarning"
+)
 models_config = ClassEnumerator(module=cuml)
 models = models_config.get_models()
 

python/cuml/tests/explainer/test_explainer_kernel_shap.py

@@ -18,6 +18,10 @@
 from cuml.testing.datasets import with_dtype
 from cuml.testing.utils import ClassEnumerator, get_shap_values
 
+# TODO(26.08): Remove this filter
+pytestmark = pytest.mark.filterwarnings(
+    "ignore:The default value of 'max_depth':FutureWarning"
+)
 models_config = ClassEnumerator(module=cuml)
 models = models_config.get_models()
 

python/cuml/tests/explainer/test_gpu_treeshap.py

@@ -27,6 +27,12 @@
 shap = pytest.importorskip("shap")
 
 
+# TODO(26.08): Remove this filter
+pytestmark = pytest.mark.filterwarnings(
+    "ignore:The default value of 'max_depth':FutureWarning"
+)
+
+
 def make_classification_with_categorical(
     *,
     n_samples,

python/cuml/tests/test_api.py

@@ -195,8 +195,10 @@ def test_mro(model):
 
 
 @pytest.mark.parametrize("model_name", list(models.keys()))
-# ignore random forest float64 warnings
+# ignore random forest float64 warnings and max_depth deprecation
 @pytest.mark.filterwarnings("ignore:To use pickling or GPU-based")
+# TODO(26.08): Remove this filter
+@pytest.mark.filterwarnings("ignore:The default value of 'max_depth'")
 def test_fit_function(dataset, model_name):
     # This test ensures that our estimators return self after a call to fit
     if model_name in [

python/cuml/tests/test_base.py

@@ -108,8 +108,10 @@ def get_param_doc(param_doc_obj, name: str):
 
 
 @pytest.mark.parametrize("child_class", list(all_base_children.keys()))
-# ignore ColumnTransformer init warning
+# ignore ColumnTransformer init warning and max_depth deprecation
 @pytest.mark.filterwarnings("ignore:Transformers are required")
+# TODO(26.08) Remove this filter
+@pytest.mark.filterwarnings("ignore:The default value of 'max_depth'")
 @pytest.mark.filterwarnings("ignore::FutureWarning")
 def test_base_children__get_param_names(child_class: str):
     """
@@ -293,6 +295,8 @@ def test_get_handle_device_ids():
         and hasattr(cls, "predict")
     ],
 )
+# TODO(26.08) Remove this filter
+@pytest.mark.filterwarnings("ignore:The default value of 'max_depth'")
 def test_regressor_predict_dtype(cls):
     X, y = make_regression(n_samples=200, random_state=42)
     X32 = X.astype("float32")
@@ -327,6 +331,8 @@ def test_regressor_predict_dtype(cls):
         (cuml.MBSGDClassifier, None),
     ],
 )
+# TODO(26.08) Remove this filter
+@pytest.mark.filterwarnings("ignore:The default value of 'max_depth'")
 @pytest.mark.parametrize(
     "target_kind", ["binary", "multiclass", "multitarget"]
 )

python/cuml/tests/test_common.py

@@ -1,5 +1,5 @@
 #
-# SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
 # SPDX-License-Identifier: Apache-2.0
 #
 import numpy as np
@@ -20,6 +20,8 @@
     ],
 )
 @pytest.mark.filterwarnings("ignore:The number of bins.*:UserWarning")
+# TODO(26.08) Remove this filter
+@pytest.mark.filterwarnings("ignore:The default value of 'max_depth'")
 def test_random_state_argument(Estimator):
     X, y = make_blobs(random_state=0)
     # Check that both integer and np.random.RandomState are accepted

python/cuml/tests/test_fil.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2019-2025, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2019-2026, NVIDIA CORPORATION.
 # SPDX-License-Identifier: Apache-2.0
 #
 
@@ -42,6 +42,11 @@
     unit_param,
 )
 
+# TODO(26.08): Remove this filter
+pytestmark = pytest.mark.filterwarnings(
+    "ignore:The default value of 'max_depth':FutureWarning"
+)
+
 
 def simulate_data(
     m,

python/cuml/tests/test_meta_estimators.py

@@ -1,5 +1,5 @@
 #
-# SPDX-FileCopyrightText: Copyright (c) 2021-2025, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2021-2026, NVIDIA CORPORATION.
 # SPDX-License-Identifier: Apache-2.0
 #
 
@@ -15,6 +15,11 @@
 from cuml.svm import SVC
 from cuml.testing.utils import ClassEnumerator
 
+# TODO(26.08): Remove this filter
+pytestmark = pytest.mark.filterwarnings(
+    "ignore:The default value of 'max_depth':FutureWarning"
+)
+
 
 def test_pipeline():
     X, y = make_classification(random_state=0)

python/cuml/tests/test_pickle.py

@@ -29,6 +29,10 @@
 )
 from cuml.tsa.arima import ARIMA
 
+# TODO(26.08) Remove this filter
+pytestmark = pytest.mark.filterwarnings(
+    "ignore:The default value of 'max_depth':FutureWarning"
+)
 regression_config = ClassEnumerator(module=cuml.linear_model)
 regression_models = regression_config.get_models()
 
@@ -369,6 +373,7 @@ def assert_model(pickled_model, X_train):
 @pytest.mark.filterwarnings(
     "ignore:Transformers((.|\n)*):UserWarning:cuml[.*]"
 )
+@pytest.mark.filterwarnings("ignore:The default value of 'max_depth'")
 def test_unfit_pickle(model_name):
     # Any model xfailed in this test cannot be used for hyperparameter sweeps
     # with dask or sklearn