[enhancement] set specific failure for host builds with "target_offload" (#2388)

* Update sycl.cpp

* Update _sycl_queue_manager.py

* Update test_config.py

* Update sycl.cpp

* Update sycl.cpp

* Update sycl.cpp

* Update sycl.cpp

* Update dal.cpp

* Update dal.cpp

* fix logic

* formatting

* convert into an object

* Update test_config.py

* Update test_config.py

* Update test_config.py

* Update test_config.py

* Update test_config.py

* Update sycl.cpp

* Update test_config.py

* Update _device_offload.py

* Update _sycl_queue_manager.py

* Update test_config.py

* Update ci.yml

* Update ci.yml

* Update sycl.cpp

* Update sycl.cpp

* Update test_config.py

* Update sycl.cpp

* Update _sycl_queue_manager.py

* Update build-and-test-lnx.yml

* Update build-and-test-lnx.yml

* Update build-and-test-lnx.yml

* Update build-and-test-lnx.yml

* Update build-and-test-lnx.yml

* Update build-and-test-lnx.yml

* Update build-and-test-lnx.yml

* Update build-and-test-lnx.yml

* Update build-and-test-lnx.yml

* Update build-and-test-lnx.yml

* Update build-and-test-lnx.yml

* Update sycl.cpp

* Update test_config.py

* Update test_config.py

* Update test_config.py

* Update sycl.cpp

* Update _config.py

Author: icfaust

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

@@ -60,6 +60,8 @@ steps:
       # 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
       pip list
     env:
       NO_DPC: ${{ variables.NO_DPC }}

onedal/_device_offload.py

@@ -30,10 +30,6 @@
 if dpctl_available:
     from dpctl.memory import MemoryUSMDevice, as_usm_memory
     from dpctl.tensor import usm_ndarray
-else:
-    from onedal import _dpc_backend
-
-    SyclQueue = getattr(_dpc_backend, "SyclQueue", None)
 
 logger = logging.getLogger("sklearnex")
 

onedal/common/sycl.cpp

@@ -21,13 +21,12 @@ namespace py = pybind11;
 
 namespace oneapi::dal::python {
 
-#ifdef ONEDAL_DATA_PARALLEL
-
 void instantiate_sycl_interfaces(py::module& m) {
     // These classes mirror a subset of functionality of the dpctl python
     // package's `SyclQueue` and `SyclDevice` objects.  In the case that dpctl
     // is not installed, these classes will enable scikit-learn-intelex to still
     // properly offload to other devices when built with the dpc backend.
+#ifdef ONEDAL_DATA_PARALLEL
     py::class_<sycl::queue> syclqueue(m, "SyclQueue");
     syclqueue.def(py::init<const sycl::device&>())
         .def(py::init([](const std::string& filter) {
@@ -81,11 +80,31 @@ void instantiate_sycl_interfaces(py::module& m) {
              })
         .def_property_readonly("is_cpu", &sycl::device::is_cpu)
         .def_property_readonly("is_gpu", &sycl::device::is_gpu);
+#else
+    struct syclqueue {};
+    py::class_<syclqueue> syclqueue(m, "SyclQueue");
+    // inspired from pybind11 PR#4698 which turns init into a no-op
+    syclqueue
+        .def(py::init([]() {
+            return nullptr;
+        }))
+        .def_static("__new__", [](const py::object& cls, const py::object& obj) {
+            // this object is defined for the host build, where SYCL support is not available.
+            // This class acts as the failure point to target_offload, which will throw an
+            // error in all circumstances if any value but the default value ("auto"), or a string
+            // starting with "cpu". The returned "queue" is a None. Must be a class to work with
+            // isinstance
+            if (!py::isinstance<py::str>(obj) || obj.cast<std::string>() != "auto") {
+                throw std::invalid_argument(
+                    "device use via `target_offload` is only supported with the DPC++ backend");
+            }
+            return py::none();
+        });
+#endif
 }
 
 ONEDAL_PY_INIT_MODULE(sycl) {
     instantiate_sycl_interfaces(m);
 }
-#endif
 
 } // namespace oneapi::dal::python

onedal/dal.cpp

@@ -43,9 +43,7 @@ ONEDAL_PY_INIT_MODULE(neighbors);
 ONEDAL_PY_INIT_MODULE(logistic_regression);
 #endif // defined(ONEDAL_VERSION) && ONEDAL_VERSION >= 20240001
 #else // ONEDAL_DATA_PARALLEL_SPMD
-#ifdef ONEDAL_DATA_PARALLEL
 ONEDAL_PY_INIT_MODULE(sycl);
-#endif // ONEDAL_DATA_PARALLEL
 
 ONEDAL_PY_INIT_MODULE(policy);
 /* datatypes*/
@@ -106,10 +104,10 @@ PYBIND11_MODULE(_onedal_py_spmd_dpc, m) {
 #else
 #ifdef ONEDAL_DATA_PARALLEL
 PYBIND11_MODULE(_onedal_py_dpc, m) {
-    init_sycl(m);
 #else
 PYBIND11_MODULE(_onedal_py_host, m) {
 #endif
+    init_sycl(m);
     init_policy(m);
     init_table(m);
 

onedal/utils/_sycl_queue_manager.py

@@ -22,9 +22,12 @@
 if dpctl_available:
     from dpctl import SyclQueue
 else:
-    from onedal import _dpc_backend
+    from onedal import _default_backend
 
-    SyclQueue = getattr(_dpc_backend, "SyclQueue", None)
+    # Use internally-defined SyclQueue defined in onedal/common/sycl.cpp
+    # the host backend SyclQueue will only accept "auto" and will return
+    # a None, it acts as a function via `__new__`. No SyclDevice is defined.
+    SyclQueue = _default_backend.SyclQueue
 
 # This special object signifies that the queue system should be
 # disabled. It will force computation to host. This occurs when the
@@ -36,12 +39,7 @@
 
 
 def __create_sycl_queue(target):
-    if SyclQueue is None:
-        # we don't have SyclQueue support
-        return None
-    if target is None:
-        return None
-    if isinstance(target, SyclQueue):
+    if isinstance(target, SyclQueue) or target is None:
         return target
     if isinstance(target, (str, int)):
         return SyclQueue(target)

sklearnex/_config.py

@@ -90,6 +90,10 @@ def set_config(
     --------
     Using ``use_raw_input=True`` is not recommended for general use as it
     bypasses data consistency checks, which may lead to unexpected behavior.
+
+    Use of ``target_offload`` requires the DPC++ backend. Setting a
+    non-default value (e.g ``cpu`` or ``gpu``) without this backend active
+    will raise an error.
     """
 
     array_api_dispatch = sklearn_configs.get("array_api_dispatch", False)

sklearnex/tests/test_config.py

@@ -16,6 +16,7 @@
 
 import logging
 
+import numpy as np
 import pytest
 import sklearn
 
@@ -127,6 +128,27 @@ def test_config_context_works():
         assert onedal_default_config_after_cc[param] == onedal_default_config[param]
 
 
+@pytest.mark.skipif(
+    onedal._default_backend.is_dpc, reason="requires host default backend"
+)
+@pytest.mark.parametrize("target", ["auto", "cpu", "cpu:0", "gpu", 3])
+def test_host_backend_target_offload(target):
+    from sklearnex.neighbors import NearestNeighbors
+
+    err_msg = (
+        r"device use via \`target_offload\` is only supported with the DPC\+\+ backend"
+    )
+
+    est = NearestNeighbors()
+    if target != "auto":
+        with pytest.raises(ValueError, match=err_msg):
+            with sklearnex.config_context(target_offload=target):
+                est.fit(np.eye(5, 8))
+    else:
+        with sklearnex.config_context(target_offload=target):
+            est.fit(np.eye(5, 8))
+
+
 @pytest.mark.skipif(
     not is_dpctl_device_available(["gpu"]), reason="Requires a gpu for fallback testing"
 )