Linear regression SPMD (#1183)

* ENH: OneDAL python ifaces for Random Forest * initial pybinds * update python wrappers, test * update init * misc tests update * update + tests * intro import * update sample weights * sklearnex dispather: cls * bump up version * update fixed params for bins * update backend wrappers * update RF cl * update onedal rf classifier fit * update fallback to base * update constraints for falback * update some attributes saving * fallback when n_outputs is not 1 * revert additions to deselected_tests.yaml * update esimators_ for cl * enabled using onedal rf regression * PEP8 fixes (#1031) * Added shape attr check in validation (#1014) Co-authored-by: Aleksandr Pivovar <[email protected]> * Fix binary AUC calculation for boolean labels (#1030) * Fix binary AUC calculation when labels are bool. Converting y_score to int using label_binarizer just like y_true. * Optional binarization for y_score Co-authored-by: Alexander Andreev <[email protected]> * Update dependency numpy [SECURITY] (#946) Co-authored-by: renovate[bot] <29139614+renovate[bot]@users.noreply.github.com> * Update dependency nbconvert to v6.5.1 [SECURITY] (#1037) Co-authored-by: renovate[bot] <29139614+renovate[bot]@users.noreply.github.com> * Update algorithm wrappers to latest Scikit-learn changes (#1032) * Update algorithm wrappers to latest sklearn changes * PEP8 and lasso fixes * Fix pca versioning * Update KMeans versioning * Rewrite kmeans.fit versions branching * Fix TSNE perplexity to pass parameter check (perplexity < n_samples) * Future warnings fix * Restore auto kmeans algorithm for compatibility with sklearn<1.1 * Numpy versions change; in-function imports removal; load_boston removal (#1041) * Bump up `dal` and `dal-devel` version to 2021.6.0 in conda recipe (#1015) * Bump up `dal` and `dal-devel` version to 2021.6.0 in conda recipe * Update meta.yaml * Code owners update (#1039) * Update of code owners * Replace DTClassifier with DTRegressor in RFRegressor (#1052) * allow n_init param in KMeans to accept str 'auto' (#1045) * allow n_init param to accept in kmeans * whitespace removed * whitespace removed * removed check for string for param n_init in kmeans * changed param 'algorithm' default to lloyd * whitespace * add versioning for param change * pep8 fix * pep8 fix * pep8 fix * Fixed typo: sklern to sklearn in README.md (#1047) * fixed typo, sklern to sklearn Co-authored-by: Alexander Andreev <[email protected]> * Update dependency jupyter-core to v4.11.2 [SECURITY] (#1056) Co-authored-by: renovate[bot] <29139614+renovate[bot]@users.noreply.github.com> * ENH: multi-gpu enable * CI updates (#1062) * Fix mpich headers for macOS * Update Ubuntu version to 22.04 * Update macos version to 12 * Change numpy version (#1057) * Change 'host' to 'cpu' in sycl_execution_content (#1064) * Change 'host' to 'cpu' in sycl_execution_content * Remove deprecated host device * Update python version to 3.10 * Revert "Update python version to 3.10" This reverts commit 9bf55d4. Co-authored-by: Samir Nasibli <[email protected]> * Fix for n_iter_ SVM attribute and oneDAL kNN classifier result option (#1071) * Add workaround for n_iter_ SVM attribute * Pytest onedal dir * Exclude test_policy * fix result_option fail for kneighbors method in knn classifier * Numpy version fix (#1070) * Downgrade build version and set test version of numpy * Upgrade version to 1.20 for python<=3.9 * Set base numpy to be 1.19.* for build * Set numpy for future python version * Enabling full versions and >=3.10 python Co-authored-by: Nikolay Petrov <[email protected]> * System info output in CI (#1073) * add system info * fix python activation on win * add system descriptor * fix descriptor calls * fix descriptor calls 2 * fix cpufeature installation on mac * change system descriptor url to oneapi-src/onedal * Add descriptor script * Remove extra variable * Update dependency certifi to v2022 [SECURITY] (#1079) Co-authored-by: renovate[bot] <29139614+renovate[bot]@users.noreply.github.com> * Fix for balanced class weight (#1080) * add balanced branch of _compute_class_weight * Remove extra computation of weights * Migration to new compiler for 2023.0 release (#1059) * link libonedal_dpc.so with libonedal_sycl.a due to mkl depency issue * only changing linker if linux * additional condition for windows linkage * removing multiply defined symbols * windows build completes, sycl errors, trying to start job * builds and 86/100 examples pass, cleaned up previous unnecessary additions * initial fix for Linux DPCPP failure * Flake8 appliance * add env variables for Linux DPCPP * Linux DPCPP ci update * Linux DPCPP check: update conda build step * moved compiler options setting into conda-recipe build * reverting linkage changes - resolved from mkl fpk update * address some comments updates passing -fsycl flag some refactoring * rebase * fix comments * Update CMakeLists.txt fix using gcc compiler * Apply pep8 * Apply pep8 n2 * Update CMakeLists.txt * resolving windows intel compiler issues * MD to MT flag for compiler and removing unneeded cmake verbose flag * updated gitignore * link libonedal_dpc.so with libonedal_sycl.a due to mkl depency issue * only changing linker if linux * additional condition for windows linkage * removing multiply defined symbols * windows build completes, sycl errors, trying to start job * builds and 86/100 examples pass, cleaned up previous unnecessary additions * initial fix for Linux DPCPP failure * Flake8 appliance * add env variables for Linux DPCPP * Linux DPCPP ci update * Linux DPCPP check: update conda build step * moved compiler options setting into conda-recipe build * reverting linkage changes - resolved from mkl fpk update * address some comments updates passing -fsycl flag some refactoring * rebase * fix comments * Update CMakeLists.txt fix using gcc compiler * Apply pep8 * Apply pep8 n2 * Update CMakeLists.txt * resolving windows intel compiler issues * MD to MT flag for compiler and removing unneeded cmake verbose flag * updated gitignore * Update setup.py * reverting MT flag back to MD * undoing last commit * reverting sycl linking addition * removing sycl addition pt. 2 * addressing Mac comments * link fsycl in cmake only when using dpc interface * FIX: daal4py spmd test * some misc fixes setup.py Co-authored-by: Samir Nasibli <[email protected]> Co-authored-by: Alexander Andreev <[email protected]> * FIX: daal4py sycl examples (#1081) Added required extra_link_args for `icpx/icx` compiler. * CI matrix and sklearn>1.1 fixes (#1086) * fix for 'normalize' deprecation in linear_model * Add initial jobs matrix * Rework matrix * Dependencies fix * Disable DPCPPROOT; fix clang-format installation * Modify sklearn dep * Modify scikit-learn installation with pip * Disable df examples/tests * fix for sed command * Fix scipy version on Windows * Fix scipy version on Windows 2 * Build docs only on py3.9 * Build docs only on py3.9 n2 * Build docs only on py3.9 n3 * Select one job for docs generation * Increase testing numpy version on Windows * Extend matrix to sklearn 1.2 * PCA n_features_ deprecation * Update linear model __init__ * Correct daal version with fixed decision forest * Move docs building to separate job * Change matrix; update deps; add sklearnex/* tests * Enable onedal iface * Temp.disable dpcpp * Update pybind11 for py311 * Add upgrade for testing deps installation * Add sourcing of MSVS on Windows * Update setup*.py * Correct requirements.txt * Correct conda-recipe/build* - d4p installation * Modify Windows testing * Fix paths for Windows testing * Fix paths for Windows testing 2 * Repair docs job * Repair docs job 2 * Repair docs job 3 * Enable older dpcpp compiler build * Try to fix deselect for daal4py/sklearn tests on Windows * fix daal4py docs build - tbb runtime libs exception * Try to fix deselect for daal4py/sklearn tests on Windows 2 * change dpcpp compiler * Change dep. versions * CI matrix change * Add global patching testing on lnx; remove extra exmples run on mac * Explicit passing of gamma paramater for kmeans (#1091) * Revert record filenames in conda-recipe (#1092) * removed outdated code * some refactoring * update onedal4py * temp: using daal4py backend for sklearnex testing leaving onlu onedal module tests * tmp: using daal4py backend * initial update for the build * misc fixes * added using spmd policy * added some python primitives for spmd policy * add instantiation for spmd interface * all together bringing updated changes from #1058 spmd interface dpctl integration updates for MPI * update onedal.spmd.ensemble * dpctl installation for ci initial * some tests fix * add missed decorator * update GPU support condition * removed merge artefacts * dev/linear-regression * All is working. Final polishing is required * Dispatching * Clang-format applied as weel as some comments from @samir-nasibli * PEP8 applied * Tests fixed * Minor * Minor * Minor * Minor * _onedal_supported fixes * Date in _common.py fixed * More date fixes * Proper initialization * Minor * PEP8 compliant * Minor * ENH: common for SPMD * PEP8 * Borrowing * Minor fixes * Minor * Comments applied * Minor * Switching by versions * Minor * Minor * More define switching * ONEDAL_STR as string * Minor * Correct DALROOT environment variable usage * No triling space * Changes in setup_*.py * Proper versions * Version define revisited * Comments by @Alexsandruss applied * Comparison in integers * Minor * Switch in linear.py * Minor * Minor * Optional testing * Correct dispatching * Minor * Bug fixes * Minor * Another try * Minor * Minor * Minor * Complete rework * PEP8 rules * Minor * Comments by @Alexsandruss * Minor * Minor * Another minor commit * Minor * Fixes for PEP8 * Minor * Fixes * New trimmed deselected list * PEP8 * Minor * Minor * Minor * Minor * Temp * Minor * Minor * Minor * Minor * Minor * Steady progress * Minor * Test for finiteness * Minor * Minor * Convertation * minor * Minor * minor * Minor * update build * pep8 codestyle * update * minor * Everything is working * License plate * Minor * Minor * Changes in testing * SPMD init changes * Minor * Minor * Further improvements * Minor * PEP8 application * Minor * Minor * required changes * Minor * Minor * CI changes * Reversed * Minor * minor * Minor improvements * Filtered list of tests * Minor changes according to @Alexsandruss * Filtering fixed * Another #ifdef --------- Co-authored-by: snasibli <[email protected]> Co-authored-by: Alexander Andreev <[email protected]> Co-authored-by: PivovarA <[email protected]> Co-authored-by: Aleksandr Pivovar <[email protected]> Co-authored-by: Furkan Tektas <[email protected]> Co-authored-by: renovate[bot] <29139614+renovate[bot]@users.noreply.github.com> Co-authored-by: msa <[email protected]> Co-authored-by: Shittu Samson <[email protected]> Co-authored-by: Nikolay Petrov <[email protected]> Co-authored-by: [email protected] <[email protected]>
uxlfoundation · Feb 25, 2023 · e5e8a13 · e5e8a13
1 parent f7fc14a
commit e5e8a13
Show file tree

Hide file tree

Showing 16 changed files with 235 additions and 15 deletions.
diff --git a/examples/sklearnex/spmd_linear_regression.py b/examples/sklearnex/spmd_linear_regression.py
@@ -0,0 +1,65 @@
+#===============================================================================
+# Copyright 2023 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#===============================================================================
+
+import numpy as np
+from mpi4py import MPI
+import dpctl
+from numpy.testing import assert_allclose
+from onedal.spmd.linear_model import LinearRegression as LinRegSpmd
+
+
+def generate_X_y(par, coef_seed, data_seed):
+    ns, nf, nr = par['ns'], par['nf'], par['nr']
+
+    crng = np.random.default_rng(coef_seed)
+    coef = crng.uniform(-4, 1, size=(nr, nf)).T
+    intp = crng.uniform(-1, 9, size=(nr, ))
+
+    drng = np.random.default_rng(data_seed)
+    data = drng.uniform(-7, 7, size=(ns, nf))
+    resp = data @ coef + intp[np.newaxis, :]
+
+    return data, resp, coef, intp
+
+
+if __name__ == "__main__":
+    q = dpctl.SyclQueue("gpu")
+
+    comm = MPI.COMM_WORLD
+    rank = comm.Get_rank()
+    size = comm.Get_size()
+
+    params_spmd = {'ns': 15, 'nf': 21, 'nr': 23}
+    params_grtr = {'ns': 77, 'nf': 21, 'nr': 23}
+
+    Xsp, ysp, csp, isp = generate_X_y(params_spmd, size, size + rank - 1)
+    Xgt, ygt, cgt, igt = generate_X_y(params_grtr, size, size + rank + 1)
+
+    assert_allclose(csp, cgt)
+    assert_allclose(isp, igt)
+
+    lrsp = LinRegSpmd(copy_X=True, fit_intercept=True)
+    lrsp.fit(Xsp, ysp, queue=q)
+
+    assert_allclose(lrsp.coef_, csp.T)
+    assert_allclose(lrsp.intercept_, isp)
+
+    ypr = lrsp.predict(Xgt, queue=q)
+
+    assert_allclose(ypr, ygt)
+
+    print("Groundtruth responses:\n", ygt)
+    print("Computed responses:\n", ypr)
diff --git a/onedal/__init__.py b/onedal/__init__.py
@@ -47,3 +47,6 @@
 
 if daal_check_version((2023, 'P', 100)):
     __all__.append('linear_model')
+
+    if _is_dpc_backend:
+        __all__.append('spmd.linear_model')
diff --git a/onedal/common/spmd_policy.cpp b/onedal/common/spmd_policy.cpp
@@ -26,7 +26,7 @@ namespace oneapi::dal::python {
 ONEDAL_PY_INIT_MODULE(spmd_policy) {
     import_dpctl();
     py::class_<dal::detail::spmd_policy<detail::data_parallel_policy>>(m, "spmd_data_parallel_policy")
-        .def(py::init([](cl::sycl::queue &q) {
+        .def(py::init([](sycl::queue &q) {
             detail::data_parallel_policy local_policy = detail::data_parallel_policy(q);
             // TODO:
             // Communicator hardcoded. Implement passing spmd communicator.

diff --git a/onedal/datatypes/__init__.py b/onedal/datatypes/__init__.py
@@ -29,8 +29,10 @@
     _get_2d_shape
 )
 
+from ._data_conversion import _convert_to_supported
+
 __all__ = ['_column_or_1d', '_validate_targets', '_check_X_y',
            '_check_array', '_check_classification_targets',
            '_type_of_target', '_is_integral_float',
            '_is_multilabel', '_check_n_features', '_num_features',
-           '_num_samples', '_get_2d_shape']
+           '_num_samples', '_get_2d_shape', '_convert_to_supported']
diff --git a/onedal/datatypes/_data_conversion.py b/onedal/datatypes/_data_conversion.py
@@ -33,3 +33,49 @@ def to_table(*args):
     if len(args) == 1:
         return convert_one_to_table(args[0])
     return (convert_one_to_table(item) for item in args)
+
+
+from onedal import _is_dpc_backend
+
+if _is_dpc_backend:
+    import numpy as np
+
+    from ..common._spmd_policy import _SPMDDataParallelInteropPolicy
+    from ..common._policy import _HostInteropPolicy, _DataParallelInteropPolicy
+
+    def _convert_to_supported_impl(policy, *data):
+        # CPUs support FP64 by default
+        is_host = isinstance(policy, _HostInteropPolicy)
+        no_dpcpp = not _is_dpc_backend
+        if is_host or no_dpcpp:
+            return data
+
+        # There is only one option of data parallel policy
+        is_dpcpp_policy = isinstance(policy, _DataParallelInteropPolicy)
+        is_spmd_policy = isinstance(policy, _SPMDDataParallelInteropPolicy)
+        assert is_spmd_policy or is_dpcpp_policy
+
+        device = policy._queue.sycl_device
+
+        def convert_or_pass(x):
+            if x.dtype is not np.float32:
+                return x.astype(np.float32)
+            else:
+                return x
+
+        if not device.has_aspect_fp64:
+            return (convert_or_pass(x) for x in data)
+        else:
+            return data
+else:
+    def _convert_to_supported_impl(policy, *data):
+        return data
+
+
+def _convert_to_supported(policy, *data):
+    res = _convert_to_supported_impl(policy, *data)
+
+    if len(data) == 1:
+        return res[0]
+    else:
+        return res
diff --git a/onedal/linear_model/linear_model.cpp b/onedal/linear_model/linear_model.cpp
@@ -204,8 +204,13 @@ ONEDAL_PY_INIT_MODULE(linear_model) {
     using task_list = types<task::regression>;
     auto sub = m.def_submodule("linear_model");
 
+#ifdef ONEDAL_DATA_PARALLEL_SPMD
+    ONEDAL_PY_INSTANTIATE(init_train_ops, sub, policy_list_spmd, task_list);
+    ONEDAL_PY_INSTANTIATE(init_infer_ops, sub, policy_list_spmd, task_list);
+#else // ONEDAL_DATA_PARALLEL_SPMD
     ONEDAL_PY_INSTANTIATE(init_train_ops, sub, policy_list, task_list);
     ONEDAL_PY_INSTANTIATE(init_infer_ops, sub, policy_list, task_list);
+#endif // ONEDAL_DATA_PARALLEL_SPMD
 
     ONEDAL_PY_INSTANTIATE(init_model, sub, task_list);
     ONEDAL_PY_INSTANTIATE(init_train_result, sub, task_list);

diff --git a/onedal/linear_model/linear_model.py b/onedal/linear_model/linear_model.py
@@ -26,7 +26,8 @@
     _num_features,
     _check_array,
     _get_2d_shape,
-    _check_n_features)
+    _check_n_features,
+    _convert_to_supported)
 
 from ..common._mixin import RegressorMixin
 from ..common._policy import _get_policy
@@ -42,6 +43,9 @@ def __init__(self, fit_intercept, copy_X, algorithm):
         self.algorithm = algorithm
         self.copy_X = copy_X
 
+    def _get_policy(self, queue, *data):
+        return _get_policy(queue, *data)
+
     def _get_onedal_params(self, dtype=np.float32):
         intercept = 'intercept|' if self.fit_intercept else ''
         return {
@@ -51,7 +55,7 @@ def _get_onedal_params(self, dtype=np.float32):
         }
 
     def _fit(self, X, y, module, queue):
-        policy = _get_policy(queue, X, y)
+        policy = self._get_policy(queue, X, y)
 
         X_loc, y_loc = X, y
         if not isinstance(X, np.ndarray):
@@ -72,6 +76,7 @@ def _fit(self, X, y, module, queue):
 
         self.n_features_in_ = _num_features(X_loc, fallback_1d=True)
 
+        X_loc, y_loc = _convert_to_supported(policy, X_loc, y_loc)
         X_table, y_table = to_table(X_loc, y_loc)
 
         result = module.train(policy, params, X_table, y_table)
@@ -111,13 +116,13 @@ def _create_model(self, module):
             assert n_targets_in == intercept.size
 
         intercept = _check_array(intercept, dtype=[np.float64, np.float32],
-                                 force_all_finite=True)
+                                 force_all_finite=True, ensure_2d=False)
         coefficients = _check_array(
             coefficients,
             dtype=[
                 np.float64,
                 np.float32],
-            force_all_finite=True)
+            force_all_finite=True, ensure_2d=False)
 
         coefficients, intercept = make2d(coefficients), make2d(intercept)
         coefficients = coefficients.T if n_targets_in == 1 else coefficients
@@ -130,16 +135,18 @@ def _create_model(self, module):
 
         packed_coefficients[:, 1:] = coefficients
         if self.fit_intercept:
-            packed_coefficients[:, :0] = intercept
+            packed_coefficients[:, 0][:, np.newaxis] = intercept
 
         m.packed_coefficients = to_table(packed_coefficients)
 
+        self._onedal_model = m
+
         return m
 
     def _predict(self, X, module, queue):
         _check_is_fitted(self)
 
-        policy = _get_policy(queue, X)
+        policy = self._get_policy(queue, X)
 
         if isinstance(X, np.ndarray):
             X_loc = np.asarray(X)
@@ -148,7 +155,7 @@ def _predict(self, X, module, queue):
 
         # Finiteness is checked in the sklearnex wrapper
         X_loc = _check_array(X_loc, dtype=[np.float64, np.float32],
-                             force_all_finite=False)
+                             force_all_finite=False, ensure_2d=False)
         _check_n_features(self, X_loc, False)
 
         params = self._get_onedal_params(X_loc)
@@ -158,7 +165,10 @@ def _predict(self, X, module, queue):
         else:
             model = self._create_model(module)
 
-        X_table = to_table(make2d(X_loc))
+        X_loc = make2d(X_loc)
+        X_loc = _convert_to_supported(policy, X_loc)
+
+        X_table = to_table(X_loc)
         result = module.infer(policy, params, model, X_table)
         y = from_table(result.responses)
 

diff --git a/onedal/linear_model/tests/test_linear_regression.py b/onedal/linear_model/tests/test_linear_regression.py
@@ -123,5 +123,6 @@ def test_reconstruct_model(queue):
         model.intercept_ = intp
 
         res = model.predict(X, queue=queue)
-
+        from onedal.datatypes._data_conversion import from_table
+        print(from_table(model._onedal_model.packed_coefficients))
         assert_allclose(gtr, res)
diff --git a/onedal/spmd/__init__.py b/onedal/spmd/__init__.py
@@ -14,4 +14,4 @@
 # limitations under the License.
 #===============================================================================
 
-__all__ = []
+__all__ = ['linear_model']
diff --git a/onedal/spmd/linear_model/__init__.py b/onedal/spmd/linear_model/__init__.py
@@ -0,0 +1,19 @@
+#===============================================================================
+# Copyright 2023 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#===============================================================================
+
+from .linear_model import LinearRegression
+
+__all__ = ['LinearRegression']
diff --git a/onedal/spmd/linear_model/linear_model.py b/onedal/spmd/linear_model/linear_model.py
@@ -0,0 +1,28 @@
+#===============================================================================
+# Copyright 2023 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#===============================================================================
+
+from abc import ABC
+from ...common._spmd_policy import _get_spmd_policy
+from onedal.linear_model import LinearRegression as LinearRegression_Batch
+
+
+class BaseLinearRegressionSPMD(ABC):
+    def _get_policy(self, queue, *data):
+        return _get_spmd_policy(queue)
+
+
+class LinearRegression(BaseLinearRegressionSPMD, LinearRegression_Batch):
+    pass
diff --git a/setup.py b/setup.py
@@ -471,7 +471,7 @@ def run(self):
         'onedal.datatypes',
         'onedal.common'
     ] + (['onedal.linear_model'] if ONEDAL_VERSION >= 20230100 else []) + (
-        ['onedal.spmd'] if build_distribute else [])),
+        ['onedal.spmd', 'onedal.spmd.linear_model'] if build_distribute else [])),
     package_data={
         'daal4py.oneapi': [
             'liboneapi_backend.so',

diff --git a/sklearnex/spmd/__init__.py b/sklearnex/spmd/__init__.py
@@ -14,4 +14,4 @@
 # limitations under the License.
 #===============================================================================
 
-__all__ = []
+__all__ = ['linear_model']
diff --git a/sklearnex/spmd/linear_model/__init__.py b/sklearnex/spmd/linear_model/__init__.py
@@ -0,0 +1,19 @@
+#===============================================================================
+# Copyright 2023 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#===============================================================================
+
+from .linear_model import LinearRegression
+
+__all__ = ['LinearRegression']
diff --git a/sklearnex/spmd/linear_model/linear_model.py b/sklearnex/spmd/linear_model/linear_model.py
@@ -0,0 +1,21 @@
+#===============================================================================
+# Copyright 2023 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#===============================================================================
+
+from onedal.spmd.linear_model import LinearRegression
+
+# TODO:
+# Currently it uses `onedal` module interface.
+# Add sklearnex dispatching.
diff --git a/tests/test_examples_sklearnex.py b/tests/test_examples_sklearnex.py
@@ -53,7 +53,8 @@ def testit(self):
     print("Generating tests for " + os.path.splitext(file)[0])
 
 
-files = [f for f in os.listdir(examples_path) if f.endswith(".py")]
+files = [f for f in os.listdir(examples_path) if f.endswith(".py") and 'spmd' not in f]
+
 for file in files:
     test_generator(file)