Merge branch 'main' into enable-security-suite

Author: gmanal

cuda_core/docs/source/release/1.2.0-notes.rst

@@ -73,6 +73,18 @@ Fixes and enhancements
   Windows, both ``ctypes.CFUNCTYPE`` and ``ctypes.WINFUNCTYPE`` are accepted.
   (`#2439 <https://github.com/NVIDIA/cuda-python/issues/2439>`__)
 
+- Starting with CUDA 13.4, unconstrained SM-resource discovery through
+  :meth:`SMResource.split` with ``SMResourceOptions(count=None)`` may return
+  every available SM, even when that count is not divisible by the device's
+  :attr:`SMResource.coscheduled_alignment`. CUDA 13.1 through 13.3 returned an
+  aligned subset for the same request. An omitted or zero
+  ``coscheduled_sm_count`` still selects the driver's default internally, but
+  CUDA 13.4 no longer guarantees that the returned :attr:`SMResource.sm_count`
+  is a multiple of that default. A green context created from the discovered
+  group may therefore span the full GPU and leave an empty remainder. Set
+  ``coscheduled_sm_count`` explicitly when an aligned result is required.
+  (`#2389 <https://github.com/NVIDIA/cuda-python/pull/2389>`__)
+
 Deprecation Notices
 -------------------
 

cuda_core/tests/test_green_context.py

@@ -405,11 +405,37 @@ def test_discovery_mode(self, sm_resource):
         assert len(groups) == 1
         assert groups[0].sm_count >= sm_resource.min_partition_size
 
-    def test_discovery_respects_alignment(self, sm_resource):
+    @pytest.mark.agent_authored(model="gpt-5.6-sol")
+    def test_by_count_discovery_respects_alignment(self, sm_resource):
+        """CUDA 12 SplitByCount discovery returns an aligned SM count."""
+        if binding_version()[0] != 12:
+            pytest.skip("test covers the CUDA 12 SplitByCount path")
+
         groups, _ = sm_resource.split(SMResourceOptions(count=None))
 
-        if sm_resource.coscheduled_alignment > 0:
-            assert groups[0].sm_count % sm_resource.coscheduled_alignment == 0
+        assert groups[0].sm_count % sm_resource.coscheduled_alignment == 0
+
+    def test_discovery_respects_explicit_coscheduled_sm_count(self, sm_resource):
+        """Constrain discovery explicitly because unconstrained discovery may use all SMs."""
+        if driver_version() < (13, 1, 0):
+            pytest.skip("explicit co-scheduled SM discovery requires CUDA 13.1+")
+
+        alignment = sm_resource.coscheduled_alignment
+        try:
+            groups, _ = sm_resource.split(
+                SMResourceOptions(
+                    count=None,
+                    coscheduled_sm_count=alignment,
+                )
+            )
+        except RuntimeError as exc:
+            pytest.skip(str(exc))
+        except CUDAError as exc:
+            if _is_invalid_resource_configuration(exc):
+                pytest.skip(str(exc))
+            raise
+
+        assert groups[0].sm_count % alignment == 0
 
     def test_two_groups(self, sm_resource):
         """Two-group split succeeds for a supported explicit request."""