FEA Add a basic RidgeCV estimator

docs/source/api/cuml.linear_model.rst

@@ -13,6 +13,7 @@ cuml.linear_model
    LinearRegression
    LogisticRegression
    Ridge
+   RidgeCV
    Lasso
    ElasticNet
    MBSGDClassifier

docs/source/api/index.rst

@@ -294,6 +294,9 @@ Regression and Classification
    * - :obj:`~cuml.linear_model.Ridge`
      - Ridge regression.
      - :mod:`cuml.linear_model`
+   * - :obj:`~cuml.linear_model.RidgeCV`
+     - Ridge regression with built-in cross-validation.
+     - :mod:`cuml.linear_model`
    * - :obj:`~cuml.solvers.SGD`
      - Stochastic Gradient Descent solver.
      - :mod:`cuml.solvers`

docs/source/cuml-accel/compatibility.rst

@@ -301,6 +301,16 @@ scores like ``sklearn.metrics.r2_score`` (for regression) or
    - If ``positive=True`` or ``solver="lbfgs"``.
 
 
+.. dropdown:: ``RidgeCV``
+   :name: ridgecv
+
+   ``RidgeCV`` will fall back to CPU in the following cases:
+
+   - If ``cv`` is not ``None`` (only the default Leave-One-Out/GCV path is
+     accelerated on GPU).
+   - If a custom ``scoring`` is provided.
+
+
 .. dropdown:: ``Lasso``
    :name: lasso
 

python/cuml/cuml/__init__.py

@@ -50,6 +50,7 @@
 from cuml.linear_model.mbsgd_classifier import MBSGDClassifier
 from cuml.linear_model.mbsgd_regressor import MBSGDRegressor
 from cuml.linear_model.ridge import Ridge
+from cuml.linear_model.ridge_cv import RidgeCV
 from cuml.manifold.t_sne import TSNE
 from cuml.manifold.umap import UMAP
 from cuml.metrics import accuracy_score, adjusted_rand_score, r2_score
@@ -155,6 +156,7 @@ def __getattr__(name):
     "RandomForestClassifier",
     "RandomForestRegressor",
     "Ridge",
+    "RidgeCV",
     "SGD",
     "SparseRandomProjection",
     "SVC",

python/cuml/cuml/accel/_overrides/sklearn/linear_model.py

@@ -18,6 +18,7 @@
     "LogisticRegression",
     "ElasticNet",
     "Ridge",
+    "RidgeCV",
     "Lasso",
 )
 
@@ -63,6 +64,15 @@ def _gpu_fit(self, X, y, sample_weight=None):
         return self
 
 
+class RidgeCV(ProxyBase):
+    _gpu_class = cuml.linear_model.RidgeCV
+
+    def _gpu_fit(self, X, y, sample_weight=None, **params):
+        # `**params` is only used by sklearn for metadata routing, which we
+        # don't support; it's absorbed here to match sklearn's signature.
+        return self._gpu.fit(X, y, sample_weight=sample_weight)
+
+
 class _ElasticNetMixin:
     def _gpu_fit(self, X, y, sample_weight=None, check_input=True):
         # check_input is ignored, only here to fix signature mismatch with sklearn

python/cuml/cuml/benchmark/algorithms.py

@@ -1,5 +1,5 @@
 #
-# SPDX-FileCopyrightText: Copyright (c) 2019-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2019-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
 import functools
@@ -389,6 +389,7 @@ def all_algorithms():
         cuml_ElasticNet = cuml.linear_model.ElasticNet
         cuml_Lasso = cuml.linear_model.Lasso
         cuml_Ridge = cuml.linear_model.Ridge
+        cuml_RidgeCV = cuml.linear_model.RidgeCV
         cuml_KernelRidge = cuml.kernel_ridge.KernelRidge
         cuml_LogisticRegression = cuml.linear_model.LogisticRegression
         cuml_RandomForestClassifier = cuml.ensemble.RandomForestClassifier
@@ -423,6 +424,7 @@ def all_algorithms():
         cuml_LinearRegression = cuml_ElasticNet = cuml_Lasso = cuml_Ridge = (
             None
         )
+        cuml_RidgeCV = None
         cuml_KernelRidge = None
         cuml_LogisticRegression = None
         cuml_RandomForestClassifier = cuml_RandomForestRegressor = None
@@ -575,6 +577,14 @@ def all_algorithms():
             accepts_labels=True,
             accuracy_function=metrics.r2_score,
         ),
+        AlgorithmPair(
+            sklearn.linear_model.RidgeCV,
+            cuml_RidgeCV,
+            shared_args={"alphas": np.logspace(-3, 3, 50)},
+            name="RidgeCV",
+            accepts_labels=True,
+            accuracy_function=metrics.r2_score,
+        ),
         AlgorithmPair(
             sklearn.kernel_ridge.KernelRidge,
             cuml_KernelRidge,

python/cuml/cuml/benchmark/automated/bench_regression.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
 #
 
@@ -64,6 +64,10 @@ def bench_ridge(gpubenchmark, bench_step, regression1):  # noqa: F811
     _benchmark_algo(gpubenchmark, "Ridge", bench_step, regression1)
 
 
+def bench_ridge_cv(gpubenchmark, bench_step, regression1):  # noqa: F811
+    _benchmark_algo(gpubenchmark, "RidgeCV", bench_step, regression1)
+
+
 def bench_knnregressor(gpubenchmark, bench_step, regression1):  # noqa: F811
     _benchmark_algo(
         gpubenchmark, "KNeighborsRegressor", bench_step, regression1

python/cuml/cuml/benchmark/configs/single_gpu.yaml

@@ -94,6 +94,30 @@ benchmarks:
           default: {rows: [2600000]}
           nightly: {rows: [10400000]}
 
+  - id: ridge_cv_fit
+    algorithm: RidgeCV
+    dataset: regression
+    operation: fit
+    tags: [linear, regression]
+    # RidgeCV runs a leave-one-out GCV over 50 alphas (O(n*p^2) per alpha), so
+    # it is much heavier than a single Ridge fit; row counts are scaled down.
+    variants:
+      narrow:
+        features: [16]
+        tiers:
+          default: {rows: [8000000]}
+          nightly: {rows: [16000000]}
+      medium:
+        features: [128]
+        tiers:
+          default: {rows: [1000000]}
+          nightly: {rows: [4000000]}
+      wide:
+        features: [512]
+        tiers:
+          default: {rows: [250000]}
+          nightly: {rows: [1000000]}
+
   - id: elasticnet_fit
     algorithm: ElasticNet
     dataset: regression

python/cuml/cuml/linear_model/__init__.py

@@ -11,3 +11,4 @@
 from cuml.linear_model.mbsgd_classifier import MBSGDClassifier
 from cuml.linear_model.mbsgd_regressor import MBSGDRegressor
 from cuml.linear_model.ridge import Ridge
+from cuml.linear_model.ridge_cv import RidgeCV