fix(cuda.core): gate cu13 managed-memory paths on driver version, not bindings only

Per PR NVIDIA#1775 review feedback (leofang). The cu13 paths in
_managed_buffer.preferred_location and _reject_numa_host_on_cuda12 gated on
binding_version() >= (13, 0, 0) only, mirroring the regressions that PR NVIDIA#2054
and PR NVIDIA#2064 fixed for other modules. With cuda-bindings 13.x installed but a
CUDA 12.x runtime driver, the v2 CU_MEM_RANGE_ATTRIBUTE_PREFERRED_LOCATION_*
attributes and the HOST_NUMA{,_CURRENT} location types fail deep in the
driver with CUDA_ERROR_INVALID_VALUE.

Both checks now AND driver_version() >= (13, 0, 0) into the gate, and the
_reject_numa_host_on_cuda12 error message names both required versions.

Author: rparolin

cuda_core/cuda/core/_memory/_managed_buffer.py

@@ -17,7 +17,7 @@
     _read_preferred_location_v2,
 )
 from cuda.core._utils.cuda_utils import driver, handle_return
-from cuda.core._utils.version import binding_version
+from cuda.core._utils.version import binding_version, driver_version
 
 if TYPE_CHECKING:
     from cuda.core._memory._buffer import MemoryResource
@@ -182,9 +182,14 @@ def preferred_location(self) -> Device | Host | None:
         host), so ``Host(numa_id=N)`` set via the setter round-trips back
         as ``Host()``.
         """
-        if binding_version() >= (13, 0, 0):
+        # The v2 path uses CU_MEM_RANGE_ATTRIBUTE_PREFERRED_LOCATION_{TYPE,ID},
+        # both added in CUDA 13. Require both bindings and the runtime driver
+        # to be 13.0+; otherwise fall back to the legacy device-ordinal path.
+        # See PR #2054 / #2064 for prior bindings-only-check regressions.
+        if binding_version() >= (13, 0, 0) and driver_version() >= (13, 0, 0):
             return _read_preferred_location_v2(self)
-        # CUDA 12 legacy path (no NUMA info available).
+        # CUDA 12 legacy path (no NUMA info available; also taken when
+        # bindings are 13.x but the runtime driver is still 12.x).
         loc_id = _get_int_attr(self, _ATTR_PREFERRED)
         if loc_id == -2:
             return None

cuda_core/cuda/core/_memory/_managed_location.py

@@ -6,7 +6,7 @@
 from dataclasses import dataclass
 from typing import Literal
 
-from cuda.core._utils.version import binding_version
+from cuda.core._utils.version import binding_version, driver_version
 
 _LocationKind = Literal["device", "host", "host_numa", "host_numa_current"]
 
@@ -34,11 +34,15 @@ def _reject_numa_host_on_cuda12(spec: _LocSpec) -> None:
     deep inside the Cython layer with ``RuntimeError``, raise a
     ``TypeError`` at the call boundary with actionable wording.
     """
-    if binding_version() >= (13, 0, 0):
+    # The host-NUMA kinds map to CU_MEM_LOCATION_TYPE_HOST_NUMA{,_CURRENT},
+    # both added in CUDA 13. Require both bindings and the runtime driver to
+    # be 13.0+; bindings-only is insufficient (PR #2054 / #2064 precedent).
+    if binding_version() >= (13, 0, 0) and driver_version() >= (13, 0, 0):
         return
     if spec.kind in ("host_numa", "host_numa_current"):
         raise TypeError(
-            "Host(numa_id=...) / Host.numa_current() require a CUDA 13 build of cuda.core; use Host() on CUDA 12"
+            "Host(numa_id=...) / Host.numa_current() require both cuda-bindings 13.0+ "
+            "and a CUDA 13+ runtime driver; use Host() instead"
         )