Remove raw inputs (#3095)

* Remove raw inputs

* formatting

* isort

* remove more instances of use_raw_inputs

---------

Co-authored-by: david-cortes-intel <david.cortes@intel.com>

Author: ethanglaser

onedal/_config.py

@@ -36,17 +36,11 @@
     If True, allows to fallback computation to sklearn after onedal
     backend in case of runtime error on onedal backend computations.
     Global default: True.
-use_raw_input:
-    If True, uses the raw input data in some SPMD onedal backend computations
-    without any checks on data consistency or validity.
-    Note: This option is not recommended for general use.
-    Global default: False.
 """
 _default_global_config = {
     "target_offload": "auto",
     "allow_fallback_to_host": False,
     "allow_sklearn_after_onedal": True,
-    "use_raw_input": False,
 }
 
 _threadlocal = threading.local()

onedal/_device_offload.py

@@ -15,21 +15,17 @@
 # ==============================================================================
 
 import inspect
-import logging
 from functools import wraps
 from operator import xor
 
 import numpy as np
 from sklearn import get_config
 
-from ._config import _get_config
 from .datatypes import copy_to_dpnp, dlpack_to_numpy
 from .utils import _sycl_queue_manager as QM
 from .utils._array_api import _asarray, _get_sycl_namespace, _is_numpy_namespace
 from .utils._third_party import is_dpnp_ndarray
 
-logger = logging.getLogger("sklearnex")
-
 
 def supports_queue(func):
     """Decorator that updates the global queue before function evaluation.
@@ -126,26 +122,7 @@ def wrapper_impl(*args, **kwargs):
         else:
             self = None
 
-        # KNeighbors*.fit can not be used with raw inputs, ignore `use_raw_input=True`
-        override_raw_input = (
-            self
-            and self.__class__.__name__ in ("KNeighborsClassifier", "KNeighborsRegressor")
-            and func.__name__ == "fit"
-            and _get_config()["use_raw_input"] is True
-        )
-        if override_raw_input:
-            pretty_name = f"{self.__class__.__name__}.{func.__name__}"
-            logger.warning(
-                f"Using raw inputs is not supported for {pretty_name}. Ignoring `use_raw_input=True` setting."
-            )
-        if _get_config()["use_raw_input"] is True and not override_raw_input:
-            if "queue" not in kwargs:
-                if usm_iface := getattr(args[0], "__sycl_usm_array_interface__", None):
-                    kwargs["queue"] = usm_iface["syclobj"]
-                else:
-                    kwargs["queue"] = None
-            return invoke_func(self, *args, **kwargs)
-        elif len(args) == 0 and len(kwargs) == 0:
+        if len(args) == 0 and len(kwargs) == 0:
             # no arguments, there's nothing we can deduce from them -> just call the function
             return invoke_func(self, *args, **kwargs)
 

sklearnex/_config.py

@@ -15,7 +15,6 @@
 # ==============================================================================
 
 import sys
-import warnings
 from contextlib import contextmanager
 
 from sklearn import get_config as skl_get_config
@@ -54,26 +53,10 @@
 {tab}
 {tab}    Global default: ``True``.
 {tab}
-{tab}use_raw_input : bool or None
-{tab}    If ``True``, uses the raw input data in some SPMD onedal backend computations
-{tab}    without any checks on data consistency or validity. Note that this can be
-{tab}    better achieved through usage of :ref:`array API classes <array_api>` without
-{tab}    ``target_offload``. Not recommended for general use.
-{tab}
-{tab}    Global default: ``False``.
-{tab}
-{tab}    .. deprecated:: 2026.0
-{tab}
 {tab}sklearn_configs : kwargs
 {tab}    Other settings accepted by scikit-learn. See :obj:`sklearn.set_config` for
 {tab}    details.
 {tab}
-{tab}Warnings
-{tab}--------
-{tab}Using ``use_raw_input=True`` is not recommended for general use as it
-{tab}bypasses data consistency checks, which may lead to unexpected behavior. It is
-{tab}recommended to use the newer :ref:`array API <array_api>` instead.
-{tab}
 {tab}Note
 {tab}----
 {tab}Usage of ``target_offload`` requires additional dependencies - see
@@ -102,7 +85,6 @@ def set_config(
     target_offload=None,
     allow_fallback_to_host=None,
     allow_sklearn_after_onedal=None,
-    use_raw_input=None,
     **sklearn_configs,
 ):  # numpydoc ignore=PR01,PR07
     """Set global configuration.
@@ -125,15 +107,6 @@ def set_config(
         local_config["allow_fallback_to_host"] = allow_fallback_to_host
     if allow_sklearn_after_onedal is not None:
         local_config["allow_sklearn_after_onedal"] = allow_sklearn_after_onedal
-    if use_raw_input is not None:
-        if use_raw_input:
-            warnings.warn(
-                "The 'use_raw_input' parameter is deprecated and will be removed in version 2026.0. "
-                "On-device input validation can now be achieved by setting 'array_api_dispatch' to True.",
-                FutureWarning,
-                stacklevel=2,
-            )
-        local_config["use_raw_input"] = use_raw_input
 
 
 set_config.__doc__ = set_config.__doc__.replace(

sklearnex/_device_offload.py

@@ -80,7 +80,7 @@ def dispatch(
 
     Depending on support conditions, oneDAL will be called, otherwise it will
     fall back to calling scikit-learn.  Dispatching to oneDAL can be influenced
-    by the 'use_raw_input' or 'allow_fallback_to_host' config parameters.
+    by the 'allow_fallback_to_host' config parameter.
 
     Parameters
     ----------
@@ -112,10 +112,6 @@ def dispatch(
         object types should match for the sklearn and onedal object methods.
     """
 
-    if get_config()["use_raw_input"]:
-        with QM.manage_global_queue(None, *args) as queue:
-            return branches["onedal"](obj, *args, **kwargs, queue=queue)
-
     # Determine if array_api dispatching is enabled, and if estimator is capable
     onedal_array_api = _array_api_offload() and get_tags(obj).onedal_array_api
     sklearn_array_api = _array_api_offload() and get_tags(obj).array_api_support
@@ -215,16 +211,8 @@ def wrapper(self, *args, **kwargs) -> Any:
             ):
                 _, (result,) = _transfer_to_host(result)
                 return result
-            # Remove check for result __sycl_usm_array_interface__ on deprecation of use_raw_inputs
-            if (
-                usm_iface := getattr(data, "__sycl_usm_array_interface__", None)
-            ) and not hasattr(result, "__sycl_usm_array_interface__"):
-                # Skip if result elements are already SYCL arrays
-                # (e.g. kneighbors tuple from from_table(like=X))
-                if isinstance(result, (tuple, list)) and all(
-                    hasattr(r, "__sycl_usm_array_interface__") for r in result
-                ):
-                    return result
+
+            if usm_iface := getattr(data, "__sycl_usm_array_interface__", None):
                 queue = usm_iface["syclobj"]
                 return copy_to_dpnp(queue, result)
 

sklearnex/basic_statistics/basic_statistics.py

@@ -21,7 +21,6 @@
 from onedal.basic_statistics import BasicStatistics as onedal_BasicStatistics
 from onedal.utils.validation import _is_csr
 
-from .._config import get_config
 from .._device_offload import dispatch
 from .._utils import PatchingConditionsChain
 from ..base import oneDALEstimator
@@ -157,20 +156,19 @@ def _onedal_gpu_supported(self, method_name, *data):
         return patching_status
 
     def _onedal_fit(self, X, sample_weight=None, queue=None):
-        if not get_config()["use_raw_input"]:
-            xp, _ = get_namespace(X, sample_weight)
-            X = validate_data(
-                self,
-                X,
-                dtype=[xp.float64, xp.float32],
-                ensure_2d=False,
-                accept_sparse="csr",
-            )
+        xp, _ = get_namespace(X, sample_weight)
+        X = validate_data(
+            self,
+            X,
+            dtype=[xp.float64, xp.float32],
+            ensure_2d=False,
+            accept_sparse="csr",
+        )
 
-            if sample_weight is not None:
-                sample_weight = _check_sample_weight(
-                    sample_weight, X, dtype=[xp.float64, xp.float32]
-                )
+        if sample_weight is not None:
+            sample_weight = _check_sample_weight(
+                sample_weight, X, dtype=[xp.float64, xp.float32]
+            )
 
         onedal_params = {
             "result_options": self.result_options,

sklearnex/basic_statistics/incremental_basic_statistics.py

@@ -23,7 +23,6 @@
     IncrementalBasicStatistics as onedal_IncrementalBasicStatistics,
 )
 
-from .._config import get_config
 from .._device_offload import dispatch
 from .._utils import PatchingConditionsChain, _add_inc_serialization_note
 from ..base import oneDALEstimator
@@ -174,8 +173,7 @@ def _onedal_finalize_fit(self, queue=None):
     def _onedal_partial_fit(self, X, sample_weight=None, queue=None, check_input=True):
         first_pass = not hasattr(self, "n_samples_seen_") or self.n_samples_seen_ == 0
 
-        # never check input when using raw input
-        if check_input and not get_config()["use_raw_input"]:
+        if check_input:
             xp, _ = get_namespace(X)
             X = validate_data(
                 self,
@@ -204,14 +202,13 @@ def _onedal_partial_fit(self, X, sample_weight=None, queue=None, check_input=Tru
         self._need_to_finalize = True
 
     def _onedal_fit(self, X, sample_weight=None, queue=None):
-        if not get_config()["use_raw_input"]:
-            xp, _ = get_namespace(X, sample_weight)
-            X = validate_data(self, X, dtype=[xp.float64, xp.float32])
+        xp, _ = get_namespace(X, sample_weight)
+        X = validate_data(self, X, dtype=[xp.float64, xp.float32])
 
-            if sample_weight is not None:
-                sample_weight = _check_sample_weight(
-                    sample_weight, X, dtype=[xp.float64, xp.float32]
-                )
+        if sample_weight is not None:
+            sample_weight = _check_sample_weight(
+                sample_weight, X, dtype=[xp.float64, xp.float32]
+            )
 
         _, n_features = X.shape
         if self.batch_size is None:

sklearnex/cluster/dbscan.py

@@ -21,7 +21,6 @@
 from onedal.cluster import DBSCAN as onedal_DBSCAN
 from onedal.utils._array_api import _is_numpy_namespace
 
-from .._config import get_config
 from .._device_offload import dispatch
 from .._utils import PatchingConditionsChain
 from ..base import oneDALEstimator
@@ -77,14 +76,11 @@ def __init__(
 
     def _onedal_fit(self, X, y, sample_weight=None, queue=None):
         xp, _ = get_namespace(X, y, sample_weight)
-        if not get_config()["use_raw_input"]:
-            X = validate_data(
-                self, X, accept_sparse="csr", dtype=[xp.float64, xp.float32]
+        X = validate_data(self, X, accept_sparse="csr", dtype=[xp.float64, xp.float32])
+        if sample_weight is not None:
+            sample_weight = _check_sample_weight(
+                sample_weight, X, dtype=[xp.float64, xp.float32]
             )
-            if sample_weight is not None:
-                sample_weight = _check_sample_weight(
-                    sample_weight, X, dtype=[xp.float64, xp.float32]
-                )
 
         onedal_params = {
             "eps": self.eps,

sklearnex/cluster/k_means.py

@@ -193,29 +193,28 @@ def _onedal_fit(self, X, _, sample_weight, queue=None):
 
             xp, _ = get_namespace(X)
 
-            if not get_config()["use_raw_input"]:
-                if _is_arraylike_not_scalar(self.init):
-                    init = validate_data(
-                        self,
-                        self.init,
-                        dtype=[xp.float64, xp.float32],
-                        accept_sparse="csr",
-                        copy=True,
-                        order="C",
-                        reset=False,
-                    )
-                    self._validate_center_shape(X, init)
-                    self.init = init
-
-                X = validate_data(
+            if _is_arraylike_not_scalar(self.init):
+                init = validate_data(
                     self,
-                    X,
-                    accept_sparse="csr",
+                    self.init,
                     dtype=[xp.float64, xp.float32],
+                    accept_sparse="csr",
+                    copy=True,
                     order="C",
-                    copy=self.copy_x,
-                    accept_large_sparse=False,
+                    reset=False,
                 )
+                self._validate_center_shape(X, init)
+                self.init = init
+
+            X = validate_data(
+                self,
+                X,
+                accept_sparse="csr",
+                dtype=[xp.float64, xp.float32],
+                order="C",
+                copy=self.copy_x,
+                accept_large_sparse=False,
+            )
 
             # Validate critical parameters to match sklearn's _check_params
             # behavior, which we bypass in the oneDAL path. This is needed
@@ -386,16 +385,13 @@ def predict(
         def _onedal_predict(self, X, sample_weight=None, queue=None):
 
             xp, _ = get_namespace(X)
-
-            if not get_config()["use_raw_input"]:
-                X = validate_data(
-                    self,
-                    X,
-                    accept_sparse="csr",
-                    reset=False,
-                    dtype=[xp.float64, xp.float32],
-                )
-
+            X = validate_data(
+                self,
+                X,
+                accept_sparse="csr",
+                reset=False,
+                dtype=[xp.float64, xp.float32],
+            )
             return self._onedal_estimator.predict(X, queue=queue)
 
         def _onedal_supported(self, method_name, *data):
@@ -456,17 +452,15 @@ def transform(self, X):
         def _onedal_transform(self, X, queue=None):
 
             xp, is_array_api = get_namespace(X)
-
-            if not get_config()["use_raw_input"]:
-                X = validate_data(
-                    self,
-                    X,
-                    accept_sparse="csr",
-                    reset=False,
-                    dtype=[xp.float64, xp.float32],
-                    order="C",
-                    accept_large_sparse=False,
-                )
+            X = validate_data(
+                self,
+                X,
+                accept_sparse="csr",
+                reset=False,
+                dtype=[xp.float64, xp.float32],
+                order="C",
+                accept_large_sparse=False,
+            )
 
             if is_array_api:
                 centers = xp.asarray(self.cluster_centers_)
@@ -500,15 +494,13 @@ def score(self, X, y=None, sample_weight=None):
         def _onedal_score(self, X, y=None, sample_weight=None, queue=None):
 
             xp, _ = get_namespace(X)
-
-            if not get_config()["use_raw_input"]:
-                X = validate_data(
-                    self,
-                    X,
-                    accept_sparse="csr",
-                    reset=False,
-                    dtype=[xp.float64, xp.float32],
-                )
+            X = validate_data(
+                self,
+                X,
+                accept_sparse="csr",
+                reset=False,
+                dtype=[xp.float64, xp.float32],
+            )
 
             if not sklearn_check_version("1.5") and sklearn_check_version("1.3"):
                 if isinstance(sample_weight, str) and sample_weight == "deprecated":