[CI] upgrade dpctl/dpnp versions

.ci/pipeline/build-and-test-lnx.yml

@@ -60,13 +60,14 @@ steps:
       pip install --upgrade -r requirements-test.txt
       pip install $(python .ci/scripts/get_compatible_scipy_version.py)
       # dpep installation is set to pypi to avoid conflict of numpy versions from pip and conda
-      # py312 is disabled due to segfault on exit of program with usage of dpctl
-      if [ $(echo $(PYTHON_VERSION) | grep '3.9\|3.10\|3.11') ] && [ $(SKLEARN_VERSION) != "1.0" ] && [ -z ${NO_DPC} ]; then pip install dpctl==0.18.* dpnp==0.16.*; fi
-      # issues exist with conda-forge dpcpp-cpp-rt=2025.1.1 it is needed to use the dpc build
-      if [ -z "${NO_DPC}" ]; then pip install dpcpp-cpp-rt==2025.1.*; fi
+      if [ $(echo $(PYTHON_VERSION) | grep '3.9\|3.10\|3.11\|3.12') ] && [ $(SKLEARN_VERSION) != "1.0" ] && [ -z ${NO_DPC} ]; then pip install dpctl==$DPCTL_VERSION dpnp==$DPNP_VERSION; fi
+      if [ -z "${NO_DPC}" ]; then pip install dpcpp-cpp-rt==$DPCPP_RT_VERSION; fi
       pip list
     env:
       NO_DPC: ${{ variables.NO_DPC }}
+      DPCTL_VERSION: 0.20.1
+      DPNP_VERSION: 0.18.1
+      DPCPP_RT_VERSION: 2025.2.0
     displayName: "Install testing requirements"
   - script: |
       . /usr/share/miniconda/etc/profile.d/conda.sh

.ci/scripts/install_dpcpp.sh

@@ -21,5 +21,5 @@ rm GPG-PUB-KEY-INTEL-SW-PRODUCTS-2023.PUB
 echo "deb https://apt.repos.intel.com/oneapi all main" | sudo tee /etc/apt/sources.list.d/oneAPI.list
 sudo add-apt-repository -y "deb https://apt.repos.intel.com/oneapi all main"
 sudo apt-get update
-sudo apt-get install -y intel-dpcpp-cpp-compiler-2025.1
+sudo apt-get install -y intel-dpcpp-cpp-compiler-2025.2
 sudo bash -c 'echo libintelocl.so > /etc/OpenCL/vendors/intel-cpu.icd'

.github/renovate.json

@@ -20,5 +20,40 @@
   ],
   "pre-commit": {
     "enabled": true
-  }
+  },
+"customManagers": [
+    {
+      "customType": "regex",
+      "fileMatch": [
+        "^\\.ci/pipeline/build-and-test-lnx\\.yml$",
+        "^\\.github/workflows/ci\\.yml$"
+      ],
+      "matchStrings": [
+        "\\s*DPCTL_VERSION: (?<currentValue>\\d+\\.\\d+\\.\\d+)"
+      ],
+      "depNameTemplate": "dpctl",
+      "datasourceTemplate": "pypi"
+    },
+    {
+      "customType": "regex",
+      "fileMatch": [
+        "^\\.ci/pipeline/build-and-test-lnx\\.yml$",
+        "^\\.github/workflows/ci\\.yml$"
+      ],
+      "matchStrings": [
+        "\\s*DPNP_VERSION: (?<currentValue>\\d+\\.\\d+\\.\\d+)"
+      ],
+      "depNameTemplate": "dpnp",
+      "datasourceTemplate": "pypi"
+    },
+    {
+      "customType": "regex",
+      "fileMatch": ["^\\.ci/pipeline/build-and-test-lnx\\.yml$"],
+      "matchStrings": [
+        "\\s*DPCPP_RT_VERSION: (?<currentValue>\\d+\\.\\d+\\.\\d+)"
+      ],
+      "depNameTemplate": "dpcpp-cpp-rt",
+      "datasourceTemplate": "pypi"
+    }
+  ]
 }

.github/workflows/ci.yml

@@ -30,8 +30,8 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  DPCTL_VERSION: 0.18.1
-  DPNP_VERSION: 0.16.0
+  DPCTL_VERSION: 0.20.1
+  DPNP_VERSION: 0.18.1
   DPCTL_PY_VERSIONS: '3.9\|3.11'
   UXL_PYTHONVERSION: "3.12"
   UXL_SKLEARNVERSION: "1.4"
@@ -463,7 +463,7 @@ jobs:
       - name: Install PyTorch
         if: contains(matrix.FRAMEWORKS, 'pytorch')
         run: |
-          pip install torch --index-url https://download.pytorch.org/whl/xpu
+          pip install torch --index-url https://download.pytorch.org/whl/cpu
           python -c "import torch; _=[print(torch.xpu.get_device_name(i)) for i in range(torch.xpu.device_count())]" 
       - name: Install daal4py/sklearnex
         run: pip install *.whl

onedal/_device_offload.py

@@ -17,14 +17,15 @@
 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, copy_to_usm, dlpack_to_numpy, usm_to_numpy
+from .datatypes import copy_to_dpnp, copy_to_usm, dlpack_to_numpy
 from .utils import _sycl_queue_manager as QM
-from .utils._array_api import _asarray, _is_numpy_namespace
+from .utils._array_api import _asarray, _get_sycl_namespace, _is_numpy_namespace
 from .utils._third_party import is_dpnp_ndarray
 
 logger = logging.getLogger("sklearnex")
@@ -62,23 +63,23 @@ def wrapper(self, *args, **kwargs):
 
 
 def _transfer_to_host(*data):
-    has_usm_data, has_host_data = False, False
+    has_usm_data = None
 
     host_data = []
     for item in data:
-        if usm_iface := getattr(item, "__sycl_usm_array_interface__", None):
-            item = usm_to_numpy(item, usm_iface)
-            has_usm_data = True
+        if item is None:
+            host_data.append(item)
+            continue
+
+        if usm_iface := hasattr(item, "__sycl_usm_array_interface__"):
+            xp = item.__array_namespace__()
+            item = xp.asnumpy(item)
+            has_usm_data = has_usm_data or has_usm_data is None
         elif not isinstance(item, np.ndarray) and (hasattr(item, "__dlpack_device__")):
             item = dlpack_to_numpy(item)
-            has_host_data = True
-        else:
-            has_host_data = True
-
-        mismatch_host_item = usm_iface is None and item is not None and has_usm_data
-        mismatch_usm_item = usm_iface is not None and has_host_data
 
-        if mismatch_host_item or mismatch_usm_item:
+        # set has_usm_data to boolean and use xor to see if they don't match
+        if xor((has_usm_data := bool(has_usm_data)), usm_iface):
             raise RuntimeError("Input data shall be located on single target device")
 
         host_data.append(item)
@@ -171,3 +172,27 @@ def wrapper_impl(*args, **kwargs):
         return result
 
     return wrapper_impl
+
+
+def support_sycl_format(func):
+    # This wrapper enables scikit-learn functions and methods to work with
+    # all sycl data frameworks as they no longer support numpy implicit
+    # conversion and must be manually converted. This is only necessary
+    # when array API is supported but not active.
+
+    @wraps(func)
+    def wrapper(*args, **kwargs):
+        if (
+            not get_config().get("array_api_dispatch", False)
+            and _get_sycl_namespace(*args)[2]
+        ):
+            with QM.manage_global_queue(kwargs.get("queue"), *args):
+                if inspect.isfunction(func) and "." in func.__qualname__:
+                    self, (args, kwargs) = args[0], _get_host_inputs(*args[1:], **kwargs)
+                    return func(self, *args, **kwargs)
+                else:
+                    args, kwargs = _get_host_inputs(*args, **kwargs)
+                    return func(*args, **kwargs)
+        return func(*args, **kwargs)
+
+    return wrapper

onedal/datatypes/__init__.py

@@ -16,7 +16,7 @@
 
 from ._data_conversion import from_table, return_type_constructor, to_table
 from ._dlpack import dlpack_to_numpy, get_torch_queue
-from ._sycl_usm import copy_to_dpnp, copy_to_usm, usm_to_numpy
+from ._sycl_usm import copy_to_dpnp, copy_to_usm
 
 __all__ = [
     "copy_to_dpnp",
@@ -26,5 +26,4 @@
     "get_torch_queue",
     "to_table",
     "return_type_constructor",
-    "usm_to_numpy",
 ]

onedal/datatypes/_data_conversion.py

@@ -95,21 +95,22 @@ def return_type_constructor(array):
         # prioritized: it provides finer-grained control of SYCL queues and the
         # related SYCL devices which are generally unavailable via DLPack
         # representations (such as SYCL contexts, SYCL sub-devices, etc.).
+        xp = array.__array_namespace__()
+        # array api support added in dpnp starting in 0.19, will fail for
+        # older versions
         if is_dpctl_tensor(array):
-            xp = array.__array_namespace__()
             func = lambda x: (
                 xp.asarray(x)
                 if hasattr(x, "__sycl_usm_array_interface__")
                 else xp.asarray(backend.from_table(x), device=device)
             )
         elif is_dpnp_ndarray(array):
-            xp = array._array_obj.__array_namespace__()
-            from_usm = array._create_from_usm_ndarray
-            func = lambda x: from_usm(
-                xp.asarray(x)
+            func = lambda x: (
+                xp.asarray(xp.as_usm_ndarray(x))
                 if hasattr(x, "__sycl_usm_array_interface__")
                 else xp.asarray(backend.from_table(x), device=device)
             )
+
     elif hasattr(array, "__array_namespace__"):
         func = array.__array_namespace__().from_dlpack
     else:

onedal/datatypes/_sycl_usm.py

@@ -17,12 +17,13 @@
 from collections.abc import Iterable
 
 import numpy as np
+import scipy.sparse as sp
 
 from ..utils._third_party import lazy_import
 
 
 @lazy_import("dpctl.memory", "dpctl.tensor")
-def array_to_usm(memory, tensor, queue, array):
+def _array_to_usm(memory, tensor, queue, array):
     try:
         mem = memory.MemoryUSMDevice(array.nbytes, queue=queue)
         mem.copy_from_host(array.tobytes())
@@ -37,42 +38,26 @@ def array_to_usm(memory, tensor, queue, array):
 
 
 @lazy_import("dpnp", "dpctl.tensor")
-def to_dpnp(dpnp, tensor, array):
+def _to_dpnp(dpnp, tensor, array):
     if isinstance(array, tensor.usm_ndarray):
         return dpnp.array(array, copy=False)
     else:
         return array
 
 
 def copy_to_usm(queue, array):
-    if hasattr(array, "__array__"):
-        return array_to_usm(queue, array)
+    if hasattr(array, "tobytes"):
+        return _array_to_usm(queue, array)
     else:
-        if isinstance(array, Iterable):
+        if isinstance(array, Iterable) and not sp.issparse(array):
             array = [copy_to_usm(queue, i) for i in array]
         return array
 
 
 def copy_to_dpnp(queue, array):
-    if hasattr(array, "__array__"):
-        return to_dpnp(array_to_usm(queue, array))
+    if hasattr(array, "tobytes"):
+        return _to_dpnp(_array_to_usm(queue, array))
     else:
-        if isinstance(array, Iterable):
+        if isinstance(array, Iterable) and not sp.issparse(array):
             array = [copy_to_dpnp(queue, i) for i in array]
         return array
-
-
-@lazy_import("dpctl.memory")
-def usm_to_numpy(memorymod, item, usm_iface):
-    buffer = memorymod.as_usm_memory(item).copy_to_host()
-    order = "C"
-    if usm_iface["strides"] is not None and len(usm_iface["strides"]) > 1:
-        if usm_iface["strides"][0] < usm_iface["strides"][1]:
-            order = "F"
-    item = np.ndarray(
-        shape=usm_iface["shape"],
-        dtype=usm_iface["typestr"],
-        buffer=buffer,
-        order=order,
-    )
-    return item

onedal/tests/test_common.py

@@ -18,6 +18,8 @@
 import os
 from glob import glob
 
+from onedal.tests.utils._dataframes_support import test_frameworks
+
 
 def _check_primitive_usage_ban(primitive_name, package, allowed_locations=None):
     """This test blocks the usage of the primitive in
@@ -55,3 +57,15 @@ def test_sklearn_check_version_ban():
     # remove this file from the list
     output = "\n".join([i for i in output if "test_common.py" not in i])
     assert output == "", f"sklearn versioning is occurring in: \n{output}"
+
+
+def test_frameworks_intentionality():
+    """Only silent skip frameworks which are not installed"""
+    fmwks = test_frameworks.replace("array_api", "array_api_strict").split(",")
+    for module in fmwks:
+        try:
+            importlib.import_module(module)
+        # If a module isn't installed, working as intended.
+        # If an ImportError occurs, then something is wrong.
+        except ModuleNotFoundError:
+            pass

onedal/utils/_array_api.py

@@ -20,6 +20,7 @@
 from functools import lru_cache
 
 import numpy as np
+import scipy.sparse as sp
 
 from ..utils._third_party import _is_subclass_fast
 
@@ -44,6 +45,8 @@ def _asarray(data, xp, *args, **kwargs):
             for i in range(len(data)):
                 result_data.append(_asarray(data[i], xp, *args, **kwargs))
             data = tuple(result_data)
+        elif sp.issparse(data):
+            pass
         else:
             for i in range(len(data)):
                 data[i] = _asarray(data[i], xp, *args, **kwargs)

sklearnex/_device_offload.py

@@ -182,8 +182,10 @@ def wrapper(self, *args, **kwargs) -> Any:
         result = func(self, *args, **kwargs)
         if not (len(args) == 0 and len(kwargs) == 0):
             data = (*args, *kwargs.values())[0]
-
-            if usm_iface := getattr(data, "__sycl_usm_array_interface__", None):
+            # 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__"):
                 queue = usm_iface["syclobj"]
                 return (
                     copy_to_dpnp(queue, result)

sklearnex/cluster/k_means.py

@@ -334,6 +334,7 @@ def _check_test_data(self, X):
             return X
 
         _transform = support_input_format(_sklearn_KMeans._transform)
+        transform = support_input_format(_sklearn_KMeans.transform)
 
         @wrap_output_data
         def score(self, X, y=None, sample_weight=None):

sklearnex/covariance/incremental_covariance.py

@@ -27,6 +27,7 @@
 
 from daal4py.sklearn._n_jobs_support import control_n_jobs
 from daal4py.sklearn._utils import daal_check_version, sklearn_check_version
+from onedal._device_offload import support_sycl_format
 from onedal.covariance import (
     IncrementalEmpiricalCovariance as onedal_IncrementalEmpiricalCovariance,
 )
@@ -231,6 +232,7 @@ def _onedal_partial_fit(self, X, queue=None, check_input=True):
         return self
 
     @wrap_output_data
+    @support_sycl_format
     def score(self, X_test, y=None):
         xp, _ = get_namespace(X_test)
 

sklearnex/decomposition/pca.py

@@ -44,6 +44,7 @@
 
     from sklearn.decomposition import PCA as _sklearn_PCA
 
+    from onedal._device_offload import support_sycl_format
     from onedal.decomposition import PCA as onedal_PCA
     from onedal.utils._array_api import _is_numpy_namespace
 
@@ -110,6 +111,10 @@ def __init__(
                 self.iterated_power = iterated_power
                 self.random_state = random_state
 
+        # guarantee operability with dpnp/dpctl, runs on CPU unless
+        # array_api_dispatch is enabled.
+        score_samples = support_sycl_format(_sklearn_PCA.score_samples)
+
         def fit(self, X, y=None):
             self._fit(X)
             return self
@@ -413,6 +418,7 @@ def _validate_n_features_in_after_fitting(self, X):
         transform.__doc__ = _sklearn_PCA.transform.__doc__
         fit_transform.__doc__ = _sklearn_PCA.fit_transform.__doc__
         inverse_transform.__doc__ = _sklearn_PCA.inverse_transform.__doc__
+        score_samples.__doc__ = _sklearn_PCA.score_samples.__doc__
 
 else:
     from daal4py.sklearn.decomposition import PCA