@@ -482,7 +482,7 @@ class Device:
     A pickled :obj:`~_device.Device` names its device by UUID. Unpickling
     returns the device with that UUID in the receiving process, whatever its
     ordinal is there, and raises ``RuntimeError`` if that device is not
-    visible.
+    visible. The pickled form is for processes on the same machine.
 
     Note
     ----
 
@@ -986,7 +986,7 @@ class Device:
     A pickled :obj:`~_device.Device` names its device by UUID. Unpickling
     returns the device with that UUID in the receiving process, whatever its
     ordinal is there, and raises ``RuntimeError`` if that device is not
-    visible.
+    visible. The pickled form is for processes on the same machine.
 
     Note
     ----
 
@@ -118,6 +118,7 @@ Fixes and enhancements
   named the device by ordinal, which names a different device when
   ``CUDA_VISIBLE_DEVICES`` differs between the processes. Unpickling raises
   ``RuntimeError`` when the device is not visible in the receiving process.
+  The pickled form is for processes on the same machine.
   (`#3196 <https://github.com/NVIDIA/cuda-python/issues/3196>`__)
 
 - Fixed a ``TypeError`` in :meth:`Program.compile` when retrying after PCH
 
@@ -5,9 +5,9 @@
 
 The child process sees the parent's two devices in reversed order, so each
 ordinal names the other device there. A ``Device`` must arrive as the device
-with the same UUID, the child must rebuild a resource for the device that
-owns the memory, and the peers of a ``VirtualMemoryResource`` must name the
-same devices as in the parent (issue #3097).
+with the same UUID (issue #3196), the child must rebuild a resource for the
+device that owns the memory, and the peers of a ``VirtualMemoryResource`` must
+name the same devices as in the parent (issue #3097).
 """
 
 import contextlib
@@ -149,15 +149,17 @@ def test_device_memory_resource_buffer(ipc_mempool_device_x2):
     """A pool-backed buffer comes back on the device that owns it, not on the device with the same ordinal."""
     dev0, dev1 = ipc_mempool_device_x2
     mr = DeviceMemoryResource(dev0, options=DeviceMemoryResourceOptions(max_size=POOL_SIZE, ipc_enabled=True))
-    buf = mr.allocate(NBYTES, stream=dev0.default_stream)
-    _fill(dev0, buf)
-
-    report = _send_to_child(buf, visible=(dev1, dev0))
-    _check_report(report, dev0)
-
-    buf.close()
-    dev0.sync()
-    mr.close()
+    buf = None
+    try:
+        buf = mr.allocate(NBYTES, stream=dev0.default_stream)
+        _fill(dev0, buf)
+        report = _send_to_child(buf, visible=(dev1, dev0))
+        _check_report(report, dev0)
+    finally:
+        if buf is not None:
+            buf.close()
+        dev0.sync()
+        mr.close()
 
 
 @pytest.mark.agent_authored(model="claude-opus-5-5")
@@ -167,14 +169,14 @@ def test_virtual_memory_resource_buffer(vmm_ipc_device_x2):
     dev0, dev1 = vmm_ipc_device_x2
     mr = _vmm_resource(dev0)
     buf = mr.allocate(NBYTES)
-    _fill(dev0, buf)
-
-    report = _send_to_child(buf, visible=(dev1, dev0))
-    _check_report(report, dev0)
-    assert report["peer_uuids"] == []
-
-    buf.close()
-    dev0.sync()
+    try:
+        _fill(dev0, buf)
+        report = _send_to_child(buf, visible=(dev1, dev0))
+        _check_report(report, dev0)
+        assert report["peer_uuids"] == []
+    finally:
+        buf.close()
+        dev0.sync()
 
 
 @pytest.mark.agent_authored(model="claude-opus-5-5")
@@ -186,15 +188,15 @@ def test_virtual_memory_resource_peers(vmm_ipc_device_x2):
         pytest.skip("Test requires peer access from the second device to the first")
     mr = _vmm_resource(dev0, peers=(dev1.device_id,))
     buf = mr.allocate(NBYTES)
-    _fill(dev0, buf)
-
-    report = _send_to_child(buf, visible=(dev1, dev0))
-    _check_report(report, dev0)
-    assert report["peer_uuids"] == [dev1.uuid]
-    assert report["peer_access"] == [3]
-
-    buf.close()
-    dev0.sync()
+    try:
+        _fill(dev0, buf)
+        report = _send_to_child(buf, visible=(dev1, dev0))
+        _check_report(report, dev0)
+        assert report["peer_uuids"] == [dev1.uuid]
+        assert report["peer_access"] == [3]
+    finally:
+        buf.close()
+        dev0.sync()
 
 
 @pytest.mark.agent_authored(model="claude-opus-5-5")
@@ -207,18 +209,18 @@ def test_virtual_memory_resource_peers_from_generator(vmm_ipc_device_x2):
     mr = _vmm_resource(dev0, peers=(peer for peer in (dev1.device_id,)))
     first = mr.allocate(NBYTES)
     buf = mr.allocate(NBYTES)
-    _fill(dev0, buf)
-    assert _access_flags(first, dev1.device_id) == 3
-    assert _access_flags(buf, dev1.device_id) == 3
-
-    report = _send_to_child(buf, visible=(dev1, dev0))
-    _check_report(report, dev0)
-    assert report["peer_uuids"] == [dev1.uuid]
-    assert report["peer_access"] == [3]
-
-    first.close()
-    buf.close()
-    dev0.sync()
+    try:
+        _fill(dev0, buf)
+        assert _access_flags(first, dev1.device_id) == 3
+        assert _access_flags(buf, dev1.device_id) == 3
+        report = _send_to_child(buf, visible=(dev1, dev0))
+        _check_report(report, dev0)
+        assert report["peer_uuids"] == [dev1.uuid]
+        assert report["peer_access"] == [3]
+    finally:
+        first.close()
+        buf.close()
+        dev0.sync()
 
 
 @pytest.mark.agent_authored(model="claude-opus-5-5")