Address review feedback on peer_accessible_by proxy

- Optimize add/discard to check the single target device (1 driver
  call) instead of scanning all devices via _query_peer_access_ids
- Add test for out-of-range int in __contains__ (returns False)
- Fix stale comment referencing removed support_multi_insert flag
- Add #2018 link to release notes entry

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: Andy-Jost

cuda_core/cuda/core/_memory/_peer_access_utils.pyx

@@ -252,15 +252,29 @@ class PeerAccessibleBySetProxy(MutableSet):
 
     def add(self, value) -> None:
         """Grant peer access from ``value`` to allocations in this pool."""
-        self._apply([value], ())
+        dev_id = _resolve_peer_device_id(value)
+        cdef DeviceMemoryResource mr = <DeviceMemoryResource>self._mr
+        if dev_id == mr._dev_id:
+            return
+        if _peer_access_includes(mr, dev_id):
+            return
+        from cuda.core._device import Device
+        if not Device(mr._dev_id).can_access_peer(dev_id):
+            raise ValueError(f"Device {mr._dev_id} cannot access peer: {dev_id}")
+        _apply_peer_access_diff(mr, (dev_id,), ())
 
     def discard(self, value) -> None:
         """Revoke peer access from ``value`` to allocations in this pool."""
         try:
             dev_id = _resolve_peer_device_id(value)
         except (TypeError, ValueError):
             return
-        self._apply((), [dev_id])
+        cdef DeviceMemoryResource mr = <DeviceMemoryResource>self._mr
+        if dev_id == mr._dev_id:
+            return
+        if not _peer_access_includes(mr, dev_id):
+            return
+        _apply_peer_access_diff(mr, (), (dev_id,))
 
     # --- bulk overrides: one driver call per op ---
 

cuda_core/docs/source/release/1.0.0-notes.rst

@@ -123,6 +123,7 @@ Breaking changes
 - :attr:`DeviceMemoryResource.peer_accessible_by` now returns a
   :class:`collections.abc.MutableSet` of :obj:`~_device.Device` objects instead
   of a sorted ``tuple[int, ...]``. The property setter is unchanged.
+  (`#2018 <https://github.com/NVIDIA/cuda-python/pull/2018>`__)
 
 - ``stream`` is now a required keyword-only argument on APIs that schedule
   work on a stream

cuda_core/tests/test_memory_peer_access.py

@@ -147,10 +147,9 @@ def test_peer_access_shared_pool_queries_driver(mempool_device_x2):
 #
 # These tests exercise the ``PeerAccessibleBySetProxy`` surface added in
 # v1.0.0. They run against ``mempool_device_x2`` because every CI machine has
-# at most 2 GPUs, which means at most one valid peer device. The relaxed
-# ``support_multi_insert=False`` path on ``assert_mutable_set_interface``
-# threads that single insertable element through the full ``MutableSet``
-# protocol.
+# at most 2 GPUs, which means at most one valid peer device. The
+# ``assert_single_member_mutable_set_interface`` helper threads that single
+# insertable element through the full ``MutableSet`` protocol.
 # ---------------------------------------------------------------------------
 
 
@@ -244,6 +243,9 @@ def test_peer_accessible_by_rejects_invalid_inputs(isolated_dmr_x2):
     # __contains__: returns False on non-coercible values, never raises
     assert "not-a-device" not in dmr.peer_accessible_by
 
+    # __contains__: out-of-range int returns False, never raises
+    assert bad_id not in dmr.peer_accessible_by
+
     # remove on a non-member raises KeyError (inherited from MutableSet)
     with pytest.raises(KeyError):
         dmr.peer_accessible_by.remove(dev1)