Merge branch 'intel:main' into dev/bs_zero

INSTALL.md

@@ -202,6 +202,14 @@ The build-process (using setup.py) happens in 4 stages:
    python setup.py develop --no-deps
    ```
 
+Where: 
+
+* Keys `--single-version-externally-managed` and `--no-deps` are required to not download daal4py after the installation of Intel(R) Extension for Scikit-learn. 
+* The `develop` mode does not install the package but creates a `.egg-link` in the deployment directory
+back to the project source-code directory. That way, you can edit the source code and see the changes
+without reinstalling the package after a small change.
+* `--single-version-externally-managed` is an option for Python packages instructing the setuptools module to create a package that the host's package manager can easily manage.
+
 - To build the python module without installing it:
 
    ```bash
@@ -217,13 +225,7 @@ python setup.py build_ext --inplace --force --abs-rpath
 python setup.py build --abs-rpath
 ```
 
-Where: 
-
-* Keys `--single-version-externally-managed` and `--no-deps` are required to not download daal4py after the installation of Intel(R) Extension for Scikit-learn. 
-* The `develop` mode does not install the package but creates a `.egg-link` in the deployment directory
-back to the project source-code directory. That way, you can edit the source code and see the changes
-without reinstalling the package after a small change.
-* `--single-version-externally-managed` is an option for Python packages instructing the setup tools module to create a package the host's package manager can easily manage.
+**Note:** when building `scikit-learn-intelex` from source with this option, it will use the oneDAL library with which it was compiled. oneDAL has dependencies on other libraries such as TBB, which is also distributed as a python package through `pip` and as a `conda` package. By default, a conda environment will first try to load TBB from its own packages if it is installed in the environment, which might cause issues if oneDAL was compiled with a system TBB instead of a conda one. In such cases, it is advised to either uninstall TBB from pip/conda (it will be loaded from the oneDAL library which links to it), or modify the order of search paths in environment variables like `${LD_LIBRARY_PATH}`.
 
 ## Build from Sources with `conda-build`
 

dependencies-dev

@@ -3,5 +3,5 @@ Jinja2==3.1.4
 numpy==2.0.1 ; python_version <= '3.9'
 numpy==2.1.3 ; python_version > '3.9'
 pybind11==2.13.6
-cmake==3.31.0.1
+cmake==3.31.1
 setuptools==75.6.0

onedal/datatypes/_data_conversion.py

@@ -35,12 +35,12 @@ def _convert_one_to_table(arg):
 def to_table(*args):
     """Create oneDAL tables from scalars and/or arrays.
 
-    Note: this implementation can be used with contiguous scipy.sparse, numpy
-    ndarrays, DPCTL/DPNP usm_ndarrays and scalars. Tables will use pointers to the
-    original array data. Scalars will be copies. Arrays may be modified in-
-    place by oneDAL during computation. This works for data located on CPU and
-    SYCL-enabled Intel GPUs. Each array may only be of a single datatype (i.e.
-    each must be homogeneous).
+    Note: this implementation can be used with scipy.sparse, numpy ndarrays,
+    DPCTL/DPNP usm_ndarrays and scalars. Tables will use pointers to the
+    original array data. Scalars and non-contiguous arrays will be copies.
+    Arrays may be modified in-place by oneDAL during computation. This works
+    for data located on CPU and SYCL-enabled Intel GPUs. Each array may only
+    be of a single datatype (i.e. each must be homogeneous).
 
     Parameters
     ----------

onedal/datatypes/data_conversion.cpp

@@ -155,18 +155,26 @@ dal::table convert_to_table(PyObject *obj) {
     }
     if (is_array(obj)) {
         PyArrayObject *ary = reinterpret_cast<PyArrayObject *>(obj);
-        if (array_is_behaved_C(ary) || array_is_behaved_F(ary)) {
+        if (!PyArray_ISCARRAY_RO(ary) && !PyArray_ISFARRAY_RO(ary)) {
+            // NOTE: this will make a C-contiguous deep copy of the data
+            // this is expected to be a special case
+            ary = PyArray_GETCONTIGUOUS(ary);
+            if (ary) {
+                res = convert_to_table(reinterpret_cast<PyObject *>(ary));
+                Py_DECREF(ary);
+                return res;
+            } 
+            else {
+                throw std::invalid_argument(
+                "[convert_to_table] Numpy input could not be converted into onedal table.");
+            }
+        }
 #define MAKE_HOMOGEN_TABLE(CType) res = convert_to_homogen_impl<CType>(ary);
-            SET_NPY_FEATURE(array_type(ary),
-                            array_type_sizeof(ary),
-                            MAKE_HOMOGEN_TABLE,
-                            throw std::invalid_argument("Found unsupported array type"));
+        SET_NPY_FEATURE(array_type(ary),
+                        array_type_sizeof(ary),
+                        MAKE_HOMOGEN_TABLE,
+                        throw std::invalid_argument("Found unsupported array type"));
 #undef MAKE_HOMOGEN_TABLE
-        }
-        else {
-            throw std::invalid_argument(
-                "[convert_to_table] Numpy input Could not convert Python object to onedal table.");
-        }
     }
     else if (strcmp(Py_TYPE(obj)->tp_name, "csr_matrix") == 0 || strcmp(Py_TYPE(obj)->tp_name, "csr_array") == 0) {
         PyObject *py_data = PyObject_GetAttrString(obj, "data");

onedal/datatypes/data_conversion_sua_iface.cpp

@@ -34,6 +34,7 @@
 
 namespace oneapi::dal::python {
 
+using namespace pybind11::literals;
 // Please follow <https://intelpython.github.io/dpctl/latest/
 // api_reference/dpctl/sycl_usm_array_interface.html#sycl-usm-array-interface-attribute>
 // for the description of `__sycl_usm_array_interface__` protocol.
@@ -42,6 +43,8 @@ namespace oneapi::dal::python {
 // of `__sycl_usm_array_interface__` protocol.
 template <typename Type>
 dal::table convert_to_homogen_impl(py::object obj) {
+    dal::table res{};
+
     // Get `__sycl_usm_array_interface__` dictionary representing USM allocations.
     auto sua_iface_dict = get_sua_interface(obj);
 
@@ -64,6 +67,25 @@ dal::table convert_to_homogen_impl(py::object obj) {
     // Get oneDAL Homogen DataLayout enumeration from input object shape and strides.
     const auto layout = get_sua_iface_layout(sua_iface_dict, r_count, c_count);
 
+    if (layout == dal::data_layout::unknown){
+        // NOTE: this will make a C-contiguous deep copy of the data
+        // if possible, this is expected to be a special case
+        py::object copy;
+        if (py::hasattr(obj, "copy")){
+            copy = obj.attr("copy")();
+        }
+        else if (py::hasattr(obj, "__array_namespace__")){
+            const auto space = obj.attr("__array_namespace__")();
+            copy = space.attr("asarray")(obj, "copy"_a = true);
+        }
+        else {
+            throw std::runtime_error("Wrong strides");
+        }
+        res = convert_to_homogen_impl<Type>(copy);
+        copy.dec_ref();
+        return res;
+    }
+
     // Get `__sycl_usm_array_interface__['data'][0]`, the first element of data entry,
     // which is a Python integer encoding USM pointer value.
     const auto* const ptr = reinterpret_cast<const Type*>(get_sua_ptr(sua_iface_dict));
@@ -79,8 +101,6 @@ dal::table convert_to_homogen_impl(py::object obj) {
     // Use read-only accessor for onedal table.
     bool is_readonly = is_sua_readonly(sua_iface_dict);
 
-    dal::table res{};
-
     if (is_readonly) {
         res = dal::homogen_table(queue,
                                  ptr,

onedal/datatypes/tests/test_data.py

@@ -377,23 +377,14 @@ def test_sua_iface_interop_unsupported_dtypes(dataframe, queue, dtype):
 def test_to_table_non_contiguous_input(dataframe, queue):
     if dataframe in "dpnp,dpctl" and not _is_dpc_backend:
         pytest.skip("__sycl_usm_array_interface__ support requires DPC backend.")
-    X = np.mgrid[:10, :10]
+    X, _ = np.mgrid[:10, :10]
     X = _convert_to_dataframe(X, sycl_queue=queue, target_df=dataframe)
     X = X[:, :3]
     sua_iface, _, _ = _get_sycl_namespace(X)
     # X expected to be non-contiguous.
     assert not X.flags.c_contiguous and not X.flags.f_contiguous
-
-    # TODO:
-    # consistent error message.
-    if dataframe in "dpnp,dpctl":
-        expected_err_msg = (
-            "Unable to convert from SUA interface: only 1D & 2D tensors are allowed"
-        )
-    else:
-        expected_err_msg = "Numpy input Could not convert Python object to onedal table."
-    with pytest.raises(ValueError, match=expected_err_msg):
-        to_table(X)
+    X_t = to_table(X)
+    assert X_t and X_t.shape == (10, 3) and X_t.has_data
 
 
 @pytest.mark.skipif(

onedal/datatypes/utils/sua_iface_helpers.cpp

@@ -163,7 +163,7 @@ dal::data_layout get_sua_iface_layout(const py::dict& sua_dict,
             return dal::data_layout::column_major;
         }
         else {
-            throw std::runtime_error("Wrong strides");
+            return dal::data_layout::unknown;
         }
     }
     else {

onedal/utils/validation.py

@@ -158,15 +158,6 @@ def _check_array(
 
     if sp.issparse(array):
         return array
-
-    # TODO: Convert this kind of arrays to a table like in daal4py
-    if not array.flags.aligned and not array.flags.writeable:
-        array = np.array(array.tolist())
-
-    # TODO: If data is not contiguous copy to contiguous
-    # Need implemeted numpy table in oneDAL
-    if not array.flags.c_contiguous and not array.flags.f_contiguous:
-        array = np.ascontiguousarray(array, array.dtype)
     return array
 
 

requirements-test.txt

@@ -9,7 +9,7 @@ numpy>=2.0.0 ; python_version >= '3.12'
 scikit-learn==1.5.2
 pandas==2.1.3 ; python_version < '3.11'
 pandas==2.2.3 ; python_version >= '3.11'
-xgboost==2.1.2
+xgboost==2.1.3
 lightgbm==4.5.0
 catboost==1.2.7 ; python_version < '3.11' # TODO: Remove 3.11 condition when catboost supports numpy 2.0
 shap==0.46.0 ; python_version < '3.13'

sklearnex/cluster/tests/test_kmeans.py

@@ -133,8 +133,6 @@ def test_dense_vs_sparse(queue, init, algorithm, dims):
     from sklearnex.cluster import KMeans
 
     if init == "random" or (not _IS_INTEL and init == "k-means++"):
-        if daal_check_version((2025, "P", 200)):
-            pytest.fail("Re-verify failure of k-means++ in 2025.2 oneDAL")
         pytest.skip(f"{init} initialization for sparse K-means is non-conformant.")
 
     # For higher level of sparsity (smaller density) the test may fail

sklearnex/tests/test_run_to_run_stability.py

@@ -153,8 +153,6 @@ def _skip_neighbors(estimator, method):
         and method
         in ["score", "predict", "kneighbors", "kneighbors_graph", "predict_proba"]
     ):
-        if daal_check_version((2025, "P", 200)):
-            pytest.fail("Re-verify failure of algorithms in oneDAL 2025.2")
         pytest.skip(f"{estimator} shows instability on non-Intel(R) hardware")