Remove redundant Python-side peer access cleanup; fix peer access tests

- Remove __dealloc__ and close() override from DeviceMemoryResource
  that cleared peer access before destruction. The C++ RAII deleter
  already handles this for owned pools (nvbug 5698116 workaround).
  For non-owned pools (default device pool), clearing peer access
  on handle disposal was incorrect behavior.

- Update peer access tests to use owned pools (DeviceMemoryResourceOptions())
  instead of default pools. Default pools are shared and may have stale
  peer access state from prior tests, causing test failures.

Made-with: Cursor

Author: Andy-Jost

cuda_core/cuda/core/_memory/_device_memory_resource.pyx

@@ -135,20 +135,6 @@ cdef class DeviceMemoryResource(_MemPool):
     def __init__(self, device_id: Device | int, options=None):
         _DMR_init(self, device_id, options)
 
-    def __dealloc__(self):
-        try:
-            self.close()
-        except Exception:
-            pass
-
-    def close(self):
-        """Close the memory resource, revoking peer access before destruction."""
-        # nvbug 5698116: clear peer access before pool destruction; also
-        # needed for non-owned (default) pools to undo modifications.
-        if self._peer_accessible_by:
-            _DMR_set_peer_accessible_by(self, [])
-        super().close()
-
     def __reduce__(self):
         return DeviceMemoryResource.from_registry, (self.uuid,)
 

cuda_core/tests/test_memory_peer_access.py

@@ -3,7 +3,7 @@
 
 import cuda.core
 import pytest
-from cuda.core import DeviceMemoryResource
+from cuda.core import DeviceMemoryResource, DeviceMemoryResourceOptions
 from cuda.core._utils.cuda_utils import CUDAError
 from helpers.buffers import PatternGen, compare_buffer_to_constant, make_scratch_buffer
 
@@ -16,7 +16,8 @@ def test_peer_access_basic(mempool_device_x2):
     zero_on_dev0 = make_scratch_buffer(dev0, 0, NBYTES)
     one_on_dev0 = make_scratch_buffer(dev0, 1, NBYTES)
     stream_on_dev0 = dev0.create_stream()
-    dmr_on_dev1 = DeviceMemoryResource(dev1)
+    # Use owned pool to ensure clean initial state (no stale peer access).
+    dmr_on_dev1 = DeviceMemoryResource(dev1, DeviceMemoryResourceOptions())
     buf_on_dev1 = dmr_on_dev1.allocate(NBYTES)
 
     # No access at first.
@@ -51,7 +52,8 @@ def test_peer_access_property_x2(mempool_device_x2):
     # The peer access list is a sorted tuple and always excludes the self
     # device.
     dev0, dev1 = mempool_device_x2
-    dmr = DeviceMemoryResource(dev0)
+    # Use owned pool to ensure clean initial state (no stale peer access).
+    dmr = DeviceMemoryResource(dev0, DeviceMemoryResourceOptions())
 
     def check(expected):
         assert isinstance(dmr.peer_accessible_by, tuple)
@@ -97,7 +99,9 @@ def test_peer_access_transitions(mempool_device_x3):
     # Allocate per-device resources.
     streams = [dev.create_stream() for dev in devs]
     pgens = [PatternGen(devs[i], NBYTES, streams[i]) for i in range(3)]
-    dmrs = [DeviceMemoryResource(dev) for dev in devs]
+    # Use owned pools (with options) to ensure clean initial state.
+    # Default pools are shared and may have stale peer access from prior tests.
+    dmrs = [DeviceMemoryResource(dev, DeviceMemoryResourceOptions()) for dev in devs]
     bufs = [dmr.allocate(NBYTES) for dmr in dmrs]
 
     def verify_state(state, pattern_seed):