Deprecate normalize option to all linear models

python/cuml/cuml/dask/common/base.py

@@ -2,6 +2,7 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 
+import warnings
 from collections.abc import Iterable
 from functools import wraps
 
@@ -470,3 +471,17 @@ def _transform_func(model, data, **kwargs):
 
 def _inverse_transform_func(model, data, **kwargs):
     return model.inverse_transform(data, **kwargs)
+
+
+def check_deprecated_normalize(model):
+    """Warn if the deprecated `normalize` option is used."""
+    if model.kwargs.get("normalize"):
+        cls_name = type(model).__name__
+        warnings.warn(
+            (
+                f"The `normalize` option to `{cls_name}` was deprecated in "
+                f"25.12 and will be removed in 26.02. Please use a `StandardScaler` "
+                f"to normalize your data external to `{cls_name}`."
+            ),
+            FutureWarning,
+        )

python/cuml/cuml/dask/linear_model/elastic_net.py

@@ -3,7 +3,7 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 
-from cuml.dask.common.base import BaseEstimator
+from cuml.dask.common.base import BaseEstimator, check_deprecated_normalize
 from cuml.dask.solvers import CD
 
 
@@ -36,10 +36,13 @@ class ElasticNet(BaseEstimator):
     fit_intercept : boolean (default = True)
         If True, Lasso tries to correct for the global mean of y.
         If False, the model expects that you have centered the data.
-    normalize : boolean (default = False)
-        If True, the predictors in X will be normalized by dividing by it's L2
-        norm.
-        If False, no scaling will be done.
+    normalize : boolean, default=False
+
+        .. deprecated:: 25.12
+            ``normalize`` is deprecated and will be removed in 26.02. When
+            needed, please use a ``StandardScaler`` to normalize your data
+            before passing to ``fit``.
+
     max_iter : int (default = 1000)
         The maximum number of iterations
     tol : float (default = 1e-3)
@@ -103,7 +106,7 @@ def fit(self, X, y):
             Dense matrix (floats or doubles) of shape (n_samples, n_features).
 
         """
-
+        check_deprecated_normalize(self)
         self.solver.fit(X, y)
         return self
 

python/cuml/cuml/dask/linear_model/lasso.py

@@ -3,7 +3,7 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 
-from cuml.dask.common.base import BaseEstimator
+from cuml.dask.common.base import BaseEstimator, check_deprecated_normalize
 from cuml.dask.solvers import CD
 
 
@@ -30,10 +30,13 @@ class Lasso(BaseEstimator):
     fit_intercept : boolean (default = True)
         If True, Lasso tries to correct for the global mean of y.
         If False, the model expects that you have centered the data.
-    normalize : boolean (default = False)
-        If True, the predictors in X will be normalized by dividing by it's L2
-        norm.
-        If False, no scaling will be done.
+    normalize : boolean, default=False
+
+        .. deprecated:: 25.12
+            ``normalize`` is deprecated and will be removed in 26.02. When
+            needed, please use a ``StandardScaler`` to normalize your data
+            before passing to ``fit``.
+
     max_iter : int (default = 1000)
         The maximum number of iterations
     tol : float (default = 1e-3)
@@ -83,6 +86,7 @@ def fit(self, X, y):
             Dense matrix (floats or doubles) of shape (n_samples, n_features).
 
         """
+        check_deprecated_normalize(self)
 
         self.solver.fit(X, y)
 

python/cuml/cuml/dask/linear_model/linear_regression.py

@@ -9,6 +9,7 @@
     BaseEstimator,
     DelayedPredictionMixin,
     SyncFitMixinLinearModel,
+    check_deprecated_normalize,
     mnmg_import,
 )
 
@@ -42,10 +43,12 @@ class LinearRegression(
         LinearRegression adds an additional term c to correct for the global
         mean of y, modeling the response as "x * beta + c".
         If False, the model expects that you have centered the data.
-    normalize : boolean (default = False)
-        If True, the predictors in X will be normalized by dividing by its
-        L2 norm.
-        If False, no scaling will be done.
+    normalize : boolean, default=False
+
+        .. deprecated:: 25.12
+            ``normalize`` is deprecated and will be removed in 26.02. When
+            needed, please use a ``StandardScaler`` to normalize your data
+            before passing to ``fit``.
 
     Attributes
     ----------
@@ -69,6 +72,7 @@ def fit(self, X, y):
         y : Dask cuDF dataframe  or CuPy backed Dask Array (n_rows, 1)
             Labels (outcome values)
         """
+        check_deprecated_normalize(self)
 
         models = self._fit(
             model_func=LinearRegression._create_model, data=(X, y)

python/cuml/cuml/dask/linear_model/ridge.py

@@ -9,6 +9,7 @@
     BaseEstimator,
     DelayedPredictionMixin,
     SyncFitMixinLinearModel,
+    check_deprecated_normalize,
     mnmg_import,
 )
 
@@ -46,10 +47,12 @@ class Ridge(BaseEstimator, SyncFitMixinLinearModel, DelayedPredictionMixin):
         If True, Ridge adds an additional term c to correct for the global
         mean of y, modeling the response as "x * beta + c".
         If False, the model expects that you have centered the data.
-    normalize : boolean (default = False)
-        If True, the predictors in X will be normalized by dividing by it's L2
-        norm.
-        If False, no scaling will be done.
+    normalize : boolean, default=False
+
+        .. deprecated:: 25.12
+            ``normalize`` is deprecated and will be removed in 26.02. When
+            needed, please use a ``StandardScaler`` to normalize your data
+            before passing to ``fit``.
 
     Attributes
     ----------
@@ -79,6 +82,7 @@ def fit(self, X, y):
         y : Dask cuDF dataframe  or CuPy backed Dask Array (n_rows, 1)
             Labels (outcome values)
         """
+        check_deprecated_normalize(self)
 
         models = self._fit(model_func=Ridge._create_model, data=(X, y))
 

python/cuml/cuml/experimental/linear_model/lars.pyx

@@ -26,6 +26,7 @@ from cuml.common.doc_utils import generate_docstring
 from cuml.internals.array import CumlArray
 from cuml.internals.base import Base
 from cuml.internals.mixins import RegressorMixin
+from cuml.linear_model.base import check_deprecated_normalize
 
 from pylibraft.common.handle cimport handle_t
 
@@ -78,14 +79,12 @@ class Lars(Base, RegressorMixin):
     fit_intercept : boolean (default = True)
         If True, Lars tries to correct for the global mean of y.
         If False, the model expects that you have centered the data.
-    normalize : boolean (default = False)
-        This parameter is ignored when `fit_intercept` is set to False.
-        If True, the predictors in X will be normalized by removing its mean
-        and dividing by it's variance. If False, then the solver expects that
-        the data is already normalized.
+    normalize : boolean, default=False
 
-        .. versionchanged:: 24.06
-            The default of `normalize` changed from `True` to `False`.
+        .. deprecated:: 25.12
+            ``normalize`` is deprecated and will be removed in 26.02. When
+            needed, please use a ``StandardScaler`` to normalize your data
+            before passing to ``fit``.
 
     copy_X : boolean (default = True)
         The solver permutes the columns of X. Set `copy_X` to True to prevent
@@ -157,7 +156,7 @@ class Lars(Base, RegressorMixin):
     coef_ = CumlArrayDescriptor()
     intercept_ = CumlArrayDescriptor()
 
-    def __init__(self, *, fit_intercept=True, normalize=True,
+    def __init__(self, *, fit_intercept=True, normalize=False,
                  handle=None, verbose=False, output_type=None, copy_X=True,
                  fit_path=True, n_nonzero_coefs=500, eps=None,
                  precompute='auto'):
@@ -291,6 +290,8 @@ class Lars(Base, RegressorMixin):
         Fit the model with X and y.
 
         """
+        check_deprecated_normalize(self)
+
         self._set_n_features_in(X)
         self._set_output_type(X)
 

python/cuml/cuml/linear_model/base.py

@@ -2,6 +2,8 @@
 # SPDX-FileCopyrightText: Copyright (c) 2020-2025, NVIDIA CORPORATION.
 # SPDX-License-Identifier: Apache-2.0
 #
+import warnings
+
 import cuml.internals
 from cuml.common.doc_utils import generate_docstring
 from cuml.internals.array import CumlArray
@@ -47,3 +49,17 @@ def predict(self, X, *, convert_dtype=True) -> CumlArray:
         out += intercept
 
         return CumlArray(out, index=X.index)
+
+
+def check_deprecated_normalize(model):
+    """Warn if the deprecated `normalize` option is used."""
+    if model.normalize:
+        cls_name = type(model).__name__
+        warnings.warn(
+            (
+                f"The `normalize` option to `{cls_name}` was deprecated in "
+                f"25.12 and will be removed in 26.02. Please use a `StandardScaler` "
+                f"to normalize your data external to `{cls_name}`."
+            ),
+            FutureWarning,
+        )

python/cuml/cuml/linear_model/elastic_net.py

@@ -13,7 +13,10 @@
     to_gpu,
 )
 from cuml.internals.mixins import FMajorInputTagMixin, RegressorMixin
-from cuml.linear_model.base import LinearPredictMixin
+from cuml.linear_model.base import (
+    LinearPredictMixin,
+    check_deprecated_normalize,
+)
 from cuml.solvers import QN
 from cuml.solvers.cd import fit_coordinate_descent
 
@@ -62,12 +65,12 @@ class ElasticNet(
         leads to significantly faster convergence especially when tol is higher
         than 1e-4.
     normalize : boolean, default=False
-        If True, the predictors in X will be normalized by dividing by the
-        column-wise standard deviation.
-        If False, no scaling will be done.
-        Note: this is in contrast to sklearn's deprecated `normalize` flag,
-        which divides by the column-wise L2 norm; but this is the same as if
-        using sklearn's StandardScaler.
+
+        .. deprecated:: 25.12
+            ``normalize`` is deprecated and will be removed in 26.02. When
+            needed, please use a ``StandardScaler`` to normalize your data
+            before passing to ``fit``.
+
     handle : cuml.Handle
         Specifies the cuml.handle that holds internal CUDA state for
         computations in this model. Most importantly, this specifies the CUDA
@@ -230,6 +233,8 @@ def fit(
         Fit the model with X and y.
 
         """
+        check_deprecated_normalize(self)
+
         if self.alpha < 0.0:
             raise ValueError(f"Expected alpha >= 0, got {self.alpha}")
         if self.selection not in ["cyclic", "random"]:

python/cuml/cuml/linear_model/lasso.py

@@ -43,13 +43,13 @@ class Lasso(ElasticNet):
         rather than looping over features sequentially by default.
         This (setting to 'random') often leads to significantly faster
         convergence especially when tol is higher than 1e-4.
-    normalize : boolean (default = False)
-        If True, the predictors in X will be normalized by dividing by the
-        column-wise standard deviation.
-        If False, no scaling will be done.
-        Note: this is in contrast to sklearn's deprecated `normalize` flag,
-        which divides by the column-wise L2 norm; but this is the same as if
-        using sklearn's StandardScaler.
+    normalize : boolean, default=False
+
+        .. deprecated:: 25.12
+            ``normalize`` is deprecated and will be removed in 26.02. When
+            needed, please use a ``StandardScaler`` to normalize your data
+            before passing to ``fit``.
+
     handle : cuml.Handle
         Specifies the cuml.handle that holds internal CUDA state for
         computations in this model. Most importantly, this specifies the CUDA

python/cuml/cuml/linear_model/linear_regression.pyx

@@ -21,7 +21,10 @@ from cuml.internals.interop import (
     to_gpu,
 )
 from cuml.internals.mixins import FMajorInputTagMixin, RegressorMixin
-from cuml.linear_model.base import LinearPredictMixin
+from cuml.linear_model.base import (
+    LinearPredictMixin,
+    check_deprecated_normalize,
+)
 
 from libc.stdint cimport uintptr_t
 from libcpp cimport bool
@@ -162,14 +165,13 @@ class LinearRegression(Base,
         If True, cuml will copy X when needed to avoid mutating the input array.
         If you're ok with X being overwritten, setting to False may avoid a copy,
         reducing memory usage for certain algorithms.
-    normalize : boolean (default = False)
-        This parameter is ignored when `fit_intercept` is set to False.
-        If True, the predictors in X will be normalized by dividing by the
-        column-wise standard deviation.
-        If False, no scaling will be done.
-        Note: this is in contrast to sklearn's deprecated `normalize` flag,
-        which divides by the column-wise L2 norm; but this is the same as if
-        using sklearn's StandardScaler.
+    normalize : boolean, default=False
+
+        .. deprecated:: 25.12
+            ``normalize`` is deprecated and will be removed in 26.02. When
+            needed, please use a ``StandardScaler`` to normalize your data
+            before passing to ``fit``.
+
     handle : cuml.Handle
         Specifies the cuml.handle that holds internal CUDA state for
         computations in this model. Most importantly, this specifies the CUDA
@@ -315,6 +317,8 @@ class LinearRegression(Base,
         Fit the model with X and y.
 
         """
+        check_deprecated_normalize(self)
+
         X_m = input_to_cuml_array(
             X,
             convert_to_dtype=(np.float32 if convert_dtype else None),

python/cuml/cuml/linear_model/linear_regression_mg.pyx

@@ -4,6 +4,7 @@
 import numpy as np
 
 import cuml.internals
+from cuml.linear_model.base import check_deprecated_normalize
 from cuml.linear_model.base_mg import MGFitMixin
 from cuml.linear_model.linear_regression import Algo, LinearRegression
 
@@ -43,6 +44,8 @@ cdef extern from "cuml/linear_model/ols_mg.hpp" namespace "ML::OLS::opg" nogil:
 class LinearRegressionMG(MGFitMixin, LinearRegression):
     @cuml.internals.api_base_return_any_skipall
     def _fit(self, X, y, coef_ptr, input_desc):
+        check_deprecated_normalize(self)
+
         cdef int algo = (
             Algo.EIG if self.algorithm == "auto" else Algo.parse(self.algorithm)
         )

python/cuml/cuml/linear_model/ridge.pyx

@@ -13,7 +13,10 @@ from cuml.internals.base import Base
 from cuml.internals.input_utils import input_to_cuml_array
 from cuml.internals.interop import InteropMixin, UnsupportedOnGPU, to_gpu
 from cuml.internals.mixins import FMajorInputTagMixin, RegressorMixin
-from cuml.linear_model.base import LinearPredictMixin
+from cuml.linear_model.base import (
+    LinearPredictMixin,
+    check_deprecated_normalize,
+)
 
 from libc.stdint cimport uintptr_t
 from libcpp cimport bool
@@ -134,13 +137,13 @@ class Ridge(Base,
     fit_intercept : boolean (default = True)
         If True, Ridge tries to correct for the global mean of y.
         If False, the model expects that you have centered the data.
-    normalize : boolean (default = False)
-        If True, the predictors in X will be normalized by dividing by the
-        column-wise standard deviation.
-        If False, no scaling will be done.
-        Note: this is in contrast to sklearn's deprecated `normalize` flag,
-        which divides by the column-wise L2 norm; but this is the same as if
-        using sklearn's StandardScaler.
+    normalize : boolean, default=False
+
+        .. deprecated:: 25.12
+            ``normalize`` is deprecated and will be removed in 26.02. When
+            needed, please use a ``StandardScaler`` to normalize your data
+            before passing to ``fit``.
+
     handle : cuml.Handle
         Specifies the cuml.handle that holds internal CUDA state for
         computations in this model. Most importantly, this specifies the CUDA
@@ -283,6 +286,8 @@ class Ridge(Base,
         """
         Fit the model with X and y.
         """
+        check_deprecated_normalize(self)
+
         cdef size_t n_rows, n_cols
         X, n_rows, n_cols, dtype = input_to_cuml_array(
             X,