Support torch validation and add to spmd tests

onedal/tests/utils/_dataframes_support.py

@@ -60,7 +60,7 @@
 from onedal.tests.utils._device_selection import get_queues
 
 test_frameworks = os.environ.get(
-    "ONEDAL_PYTEST_FRAMEWORKS", "numpy,pandas,dpnp,array_api"
+    "ONEDAL_PYTEST_FRAMEWORKS", "numpy,pandas,dpnp,array_api,torch"
 )
 
 
@@ -126,6 +126,8 @@ def get_df_and_q(dataframe: str):
         or array_api_enabled()
     ):
         dataframes_and_queues.append(pytest.param("array_api", None, id="array_api"))
+    if torch_available and "torch" in dataframe_filter_:
+        dataframes_and_queues.extend(get_df_and_q("torch"))
 
     return dataframes_and_queues
 
@@ -134,6 +136,8 @@ def _as_numpy(obj, *args, **kwargs):
     """Converted input object to numpy.ndarray format."""
     if dpnp_available and isinstance(obj, dpnp.ndarray):
         return obj.asnumpy(*args, **kwargs)
+    if torch_available and isinstance(obj, torch.Tensor):
+        return obj.cpu().detach().numpy(*args, **kwargs)
     if isinstance(obj, pd.DataFrame) or isinstance(obj, pd.Series):
         return obj.to_numpy(*args, **kwargs)
     if sp.issparse(obj):
@@ -172,5 +176,18 @@ def _convert_to_dataframe(obj, sycl_queue=None, target_df=None, *args, **kwargs)
 
         xp = array_api_modules[target_df]
         return xp.asarray(obj)
+    elif target_df == "torch":
+        if "dtype" in kwargs:
+            kwargs["dtype"] = torch.from_numpy(np.empty(0, dtype=kwargs["dtype"])).dtype
+        # Mirror the requested sycl_queue's device so torch tensors don't land on
+        # xpu for CPU-queue cases (dpnp honors sycl_queue; torch must too).
+        is_gpu = sycl_queue is not None and getattr(
+            sycl_queue.sycl_device, "is_gpu", False
+        )
+        if is_gpu and hasattr(torch, "xpu") and torch.xpu.is_available():
+            device = "xpu"
+        else:
+            device = "cpu"
+        return torch.as_tensor(obj, device=device, *args, **kwargs)
 
     raise RuntimeError("Unsupported dataframe conversion")

onedal/utils/_array_api.py

@@ -22,7 +22,7 @@
 import numpy as np
 import scipy.sparse as sp
 
-from ..utils._third_party import _is_subclass_fast
+from ..utils._third_party import _is_subclass_fast, is_torch_tensor, lazy_import
 
 
 def _supports_buffer_protocol(obj):
@@ -73,17 +73,35 @@ def _cls_to_sycl_namespace(cls):
         raise ValueError(f"SYCL type not recognized: {cls}")
 
 
+@lazy_import("array_api_compat")
+def _torch_namespace(array_api_compat, array):
+    # torch's array API namespace is exposed via array_api_compat, not as a
+    # __array_namespace__ attribute on the tensor itself.
+    return array_api_compat.get_namespace(array)
+
+
+def _is_sycl_array(x):
+    # dpnp exposes __sycl_usm_array_interface__; torch xpu tensors do not, so
+    # they are detected separately. Both must be recognized regardless of the
+    # array_api_dispatch global so compute-follows-data is preserved.
+    return hasattr(x, "__sycl_usm_array_interface__") or is_torch_tensor(x)
+
+
 def _get_sycl_namespace(*arrays):
     """Get namespace of sycl arrays."""
 
     # sycl support designed to work regardless of array_api_dispatch sklearn global value
-    sua_iface = {type(x): x for x in arrays if hasattr(x, "__sycl_usm_array_interface__")}
+    sua_iface = {type(x): x for x in arrays if _is_sycl_array(x)}
 
     if len(sua_iface) > 1:
         raise ValueError(f"Multiple SYCL types for array inputs: {sua_iface}")
 
     if sua_iface:
         (X,) = sua_iface.values()
+        if is_torch_tensor(X):
+            # torch is array-API compliant via array_api_compat; report it as
+            # such so downstream conversions keep results as torch tensors.
+            return sua_iface, _torch_namespace(X), True
         return (
             sua_iface,
             _cls_to_sycl_namespace(type(X)),

sklearnex/spmd/basic_statistics/tests/test_basic_statistics_spmd.py

@@ -38,7 +38,7 @@
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.mpi
 def test_basic_stats_spmd_gold(dataframe, queue):
@@ -82,7 +82,7 @@ def test_basic_stats_spmd_gold(dataframe, queue):
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("array_api_dispatch", [True, False])
 @pytest.mark.mpi

sklearnex/spmd/basic_statistics/tests/test_incremental_basic_statistics_spmd.py

@@ -38,7 +38,7 @@
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("weighted", [True, False])
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
@@ -104,7 +104,7 @@ def test_incremental_basic_statistics_fit_spmd_gold(dataframe, queue, weighted,
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("num_blocks", [1, 2])
 @pytest.mark.parametrize("weighted", [True, False])
@@ -178,7 +178,7 @@ def test_incremental_basic_statistics_partial_fit_spmd_gold(
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("num_blocks", [1, 2])
 @pytest.mark.parametrize("weighted", [True, False])
@@ -247,7 +247,7 @@ def test_incremental_basic_statistics_single_option_partial_fit_spmd_gold(
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("num_blocks", [1, 2])
 @pytest.mark.parametrize("weighted", [True, False])

sklearnex/spmd/cluster/tests/test_dbscan_spmd.py

@@ -36,7 +36,7 @@
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.mpi
 def test_dbscan_spmd_gold(dataframe, queue):
@@ -67,7 +67,7 @@ def test_dbscan_spmd_gold(dataframe, queue):
 @pytest.mark.parametrize("min_samples", [2, 5, 15])
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
 @pytest.mark.parametrize("array_api_dispatch", [True, False])

sklearnex/spmd/cluster/tests/test_kmeans_spmd.py

@@ -39,7 +39,7 @@
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.mpi
 def test_kmeans_spmd_gold(dataframe, queue):
@@ -107,7 +107,7 @@ def test_kmeans_spmd_gold(dataframe, queue):
 @pytest.mark.parametrize("n_clusters", [2, 5, 15])
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
 @pytest.mark.parametrize("array_api_dispatch", [True, False])

sklearnex/spmd/covariance/tests/test_covariance_spmd.py

@@ -37,7 +37,7 @@
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.mpi
 def test_covariance_spmd_gold(dataframe, queue):
@@ -83,7 +83,7 @@ def test_covariance_spmd_gold(dataframe, queue):
 @pytest.mark.parametrize("assume_centered", [True, False])
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
 @pytest.mark.parametrize("array_api_dispatch", [True, False])

sklearnex/spmd/covariance/tests/test_incremental_covariance_spmd.py

@@ -37,7 +37,7 @@
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("assume_centered", [True, False])
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
@@ -88,7 +88,7 @@ def test_incremental_covariance_fit_spmd_gold(dataframe, queue, assume_centered,
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("num_blocks", [1, 2])
 @pytest.mark.parametrize("assume_centered", [True, False])
@@ -149,7 +149,7 @@ def test_incremental_covariance_partial_fit_spmd_gold(
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("array_api_dispatch", [True, False])
 @pytest.mark.mpi

sklearnex/spmd/decomposition/tests/test_incremental_pca_spmd.py

@@ -47,7 +47,7 @@
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("whiten", [True, False])
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
@@ -95,7 +95,7 @@ def test_incremental_pca_fit_spmd_gold(dataframe, queue, whiten, dtype):
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("whiten", [True, False])
 @pytest.mark.parametrize("num_blocks", [1, 2])
@@ -159,7 +159,7 @@ def test_incremental_pca_partial_fit_spmd_gold(
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("whiten", [True, False])
 @pytest.mark.parametrize("n_components", [None, 2, 5])
@@ -211,7 +211,7 @@ def test_incremental_pca_fit_spmd_random(
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("whiten", [True, False])
 @pytest.mark.parametrize("n_components", [None, 2, 5])

sklearnex/spmd/decomposition/tests/test_pca_spmd.py

@@ -37,7 +37,7 @@
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.mpi
 def test_pca_spmd_gold(dataframe, queue):
@@ -90,7 +90,7 @@ def test_pca_spmd_gold(dataframe, queue):
 @pytest.mark.parametrize("whiten", [True, False])
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
 @pytest.mark.parametrize("array_api_dispatch", [True, False])

sklearnex/spmd/ensemble/tests/test_forest_spmd.py

@@ -38,7 +38,7 @@
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.mpi
 def test_rfcls_spmd_gold(dataframe, queue):
@@ -107,7 +107,7 @@ def test_rfcls_spmd_gold(dataframe, queue):
 @pytest.mark.parametrize("local_trees_mode", [False, True])
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
 @pytest.mark.parametrize("array_api_dispatch", [True, False])
@@ -173,7 +173,7 @@ def test_rfcls_spmd_synthetic(
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.mpi
 def test_rfreg_spmd_gold(dataframe, queue):
@@ -242,7 +242,7 @@ def test_rfreg_spmd_gold(dataframe, queue):
 @pytest.mark.parametrize("local_trees_mode", [False, True])
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
 @pytest.mark.parametrize("array_api_dispatch", [True, False])

sklearnex/spmd/linear_model/tests/test_incremental_linear_spmd.py

@@ -37,7 +37,7 @@
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("fit_intercept", [True, False])
 @pytest.mark.parametrize("macro_block", [None, 1024])
@@ -112,7 +112,7 @@ def test_incremental_linear_regression_fit_spmd_gold(
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("fit_intercept", [True, False])
 @pytest.mark.parametrize("num_blocks", [1, 2])
@@ -196,7 +196,7 @@ def test_incremental_linear_regression_partial_fit_spmd_gold(
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("fit_intercept", [True, False])
 @pytest.mark.parametrize("num_samples", [100, 1000])
@@ -261,7 +261,7 @@ def test_incremental_linear_regression_fit_spmd_random(
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("fit_intercept", [True, False])
 @pytest.mark.parametrize("num_blocks", [1, 2])

sklearnex/spmd/linear_model/tests/test_linear_regression_spmd.py

@@ -38,7 +38,7 @@
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.mpi
 def test_linear_spmd_gold(dataframe, queue):
@@ -102,7 +102,7 @@ def test_linear_spmd_gold(dataframe, queue):
 @pytest.mark.parametrize("n_features", [10, 100])
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
 @pytest.mark.parametrize("array_api_dispatch", [True, False])

sklearnex/spmd/linear_model/tests/test_logistic_regression_spmd.py

@@ -38,7 +38,7 @@
 )
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.mpi
 def test_logistic_spmd_gold(dataframe, queue):
@@ -112,7 +112,7 @@ def test_logistic_spmd_gold(dataframe, queue):
 @pytest.mark.parametrize("tol", [1e-2, 1e-4])
 @pytest.mark.parametrize(
     "dataframe,queue",
-    get_dataframes_and_queues(dataframe_filter_="dpnp", device_filter_="gpu"),
+    get_dataframes_and_queues(dataframe_filter_="dpnp,torch", device_filter_="gpu"),
 )
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
 @pytest.mark.parametrize("array_api_dispatch", [True, False])