diff --git a/cuda_bindings/tests/test_cufile.py b/cuda_bindings/tests/test_cufile.py
index 46bd8429a62..7de4ceb2e20 100644
--- a/cuda_bindings/tests/test_cufile.py
+++ b/cuda_bindings/tests/test_cufile.py
@@ -138,7 +138,7 @@ def ctx():
     (err,) = cuda.cuCtxSetCurrent(ctx)
     assert err == cuda.CUresult.CUDA_SUCCESS
 
-    yield
+    yield ctx
 
     cuda.cuDevicePrimaryCtxRelease(device)
 
diff --git a/cuda_bindings/tests/test_examples.py b/cuda_bindings/tests/test_examples.py
index 652515830f8..a7a0524a217 100644
--- a/cuda_bindings/tests/test_examples.py
+++ b/cuda_bindings/tests/test_examples.py
@@ -19,6 +19,7 @@ def test_example(example):
 
     env = os.environ.copy()
     env["CUDA_BINDINGS_SKIP_EXAMPLE"] = "100"
+    env["MPLBACKEND"] = "Agg"  # avoid plt.show() from blocking
 
     process = subprocess.run([sys.executable, example], capture_output=True, env=env)  # noqa: S603
     # returncode is a special value used in the examples to indicate that system requirements are not met.
diff --git a/cuda_core/tests/graph/test_graph_definition.py b/cuda_core/tests/graph/test_graph_definition.py
index 4b3122c763e..f8f427567a0 100644
--- a/cuda_core/tests/graph/test_graph_definition.py
+++ b/cuda_core/tests/graph/test_graph_definition.py
@@ -577,20 +577,6 @@ def node_spec(request, init_cuda):
 # =============================================================================
 
 
-@pytest.fixture
-def sample_graphdef(init_cuda):
-    """A sample GraphDefinition for standalone tests."""
-    return GraphDefinition()
-
-
-@pytest.fixture
-def dot_file(tmp_path):
-    """Temporary DOT file path, cleaned up after test."""
-    path = tmp_path / "graph.dot"
-    yield path
-    path.unlink(missing_ok=True)
-
-
 # =============================================================================
 # Topology tests (parameterized over graph specs)
 # =============================================================================
@@ -830,14 +816,16 @@ def registered(node):
 # =============================================================================
 
 
-def test_graphdef_handle_valid(sample_graphdef):
+def test_graphdef_handle_valid(init_cuda):
     """GraphDefinition has a valid non-null handle."""
+    sample_graphdef = GraphDefinition()
     assert sample_graphdef.handle is not None
     assert int(sample_graphdef.handle) != 0
 
 
-def test_graphdef_entry_is_virtual(sample_graphdef):
+def test_graphdef_entry_is_virtual(init_cuda):
     """Internal entry node is virtual (no pred/succ, type is None)."""
+    sample_graphdef = GraphDefinition()
     entry = sample_graphdef._entry
     assert isinstance(entry, GraphNode)
     assert entry.pred == set()
@@ -850,8 +838,9 @@ def test_graphdef_entry_is_virtual(sample_graphdef):
 # =============================================================================
 
 
-def test_alloc_zero_size_fails(sample_graphdef):
+def test_alloc_zero_size_fails(init_cuda):
     """Alloc with zero size raises error (CUDA limitation)."""
+    sample_graphdef = GraphDefinition()
     _skip_if_no_mempool()
     from cuda.core._utils.cuda_utils import CUDAError
 
@@ -859,8 +848,9 @@ def test_alloc_zero_size_fails(sample_graphdef):
         sample_graphdef.allocate(0)
 
 
-def test_free_creates_dependency(sample_graphdef):
+def test_free_creates_dependency(init_cuda):
     """Free node depends on its predecessor."""
+    sample_graphdef = GraphDefinition()
     _skip_if_no_mempool()
     with xfail_on_graph_mempool_oom():
         alloc = sample_graphdef.allocate(ALLOC_SIZE)
@@ -868,8 +858,9 @@ def test_free_creates_dependency(sample_graphdef):
     assert alloc in free.pred
 
 
-def test_alloc_free_chain(sample_graphdef):
+def test_alloc_free_chain(init_cuda):
     """Alloc and free can be chained."""
+    sample_graphdef = GraphDefinition()
     _skip_if_no_mempool()
     with xfail_on_graph_mempool_oom():
         a1 = sample_graphdef.allocate(ALLOC_SIZE)
@@ -886,8 +877,9 @@ def test_alloc_free_chain(sample_graphdef):
 # =============================================================================
 
 
-def test_alloc_memory_type_invalid(sample_graphdef):
+def test_alloc_memory_type_invalid(init_cuda):
     """Invalid memory type raises ValueError."""
+    sample_graphdef = GraphDefinition()
     with pytest.raises(ValueError, match="'invalid' is not a valid GraphMemoryType. Must be "):
         sample_graphdef.allocate(ALLOC_SIZE, memory_type="invalid")
 
@@ -899,8 +891,9 @@ def test_alloc_memory_type_invalid(sample_graphdef):
         pytest.param(lambda d: d, id="Device_object"),
     ],
 )
-def test_alloc_device_option(sample_graphdef, device_spec):
+def test_alloc_device_option(init_cuda, device_spec):
     """Device can be specified as int or Device object."""
+    sample_graphdef = GraphDefinition()
     _skip_if_no_mempool()
     device = Device()
     with xfail_on_graph_mempool_oom(device):
@@ -923,8 +916,9 @@ def test_alloc_peer_access(mempool_device_x2):
 
 
 @pytest.mark.parametrize("num_branches", [2, 3, 5])
-def test_join_merges_branches(sample_graphdef, num_branches):
+def test_join_merges_branches(init_cuda, num_branches):
     """join() with multiple branches creates correct dependencies."""
+    sample_graphdef = GraphDefinition()
     _skip_if_no_mempool()
     with xfail_on_graph_mempool_oom():
         branches = [sample_graphdef.allocate(ALLOC_SIZE) for _ in range(num_branches)]
@@ -938,8 +932,9 @@ def test_join_merges_branches(sample_graphdef, num_branches):
 # =============================================================================
 
 
-def test_launch_creates_node(sample_graphdef):
+def test_launch_creates_node(init_cuda):
     """launch() creates a KernelNode."""
+    sample_graphdef = GraphDefinition()
     mod = compile_common_kernels()
     kernel = mod.get_kernel("empty_kernel")
     config = LaunchConfig(grid=1, block=1)
@@ -947,8 +942,9 @@ def test_launch_creates_node(sample_graphdef):
     assert isinstance(node, KernelNode)
 
 
-def test_launch_chain_dependencies(sample_graphdef):
+def test_launch_chain_dependencies(init_cuda):
     """Chained launches create correct dependencies."""
+    sample_graphdef = GraphDefinition()
     mod = compile_common_kernels()
     kernel = mod.get_kernel("empty_kernel")
     config = LaunchConfig(grid=1, block=1)
@@ -1010,15 +1006,17 @@ def _instantiate_and_upload(graph_definition, kwargs, stream):
 
 
 @pytest.mark.parametrize("inst_kwargs", _INSTANTIATE_ONLY_OPTIONS)
-def test_instantiate_empty_graph(sample_graphdef, inst_kwargs):
+def test_instantiate_empty_graph(init_cuda, inst_kwargs):
     """Empty graph can be instantiated."""
+    sample_graphdef = GraphDefinition()
     graph = _instantiate(sample_graphdef, inst_kwargs)
     assert graph is not None
 
 
 @pytest.mark.parametrize("inst_kwargs", _INSTANTIATE_ONLY_OPTIONS)
-def test_instantiate_with_nodes(sample_graphdef, inst_kwargs):
+def test_instantiate_with_nodes(init_cuda, inst_kwargs):
     """Graph with nodes can be instantiated."""
+    sample_graphdef = GraphDefinition()
     _skip_if_no_mempool()
     with xfail_on_graph_mempool_oom():
         sample_graphdef.allocate(ALLOC_SIZE)
@@ -1028,8 +1026,9 @@ def test_instantiate_with_nodes(sample_graphdef, inst_kwargs):
 
 
 @pytest.mark.skipif(not Device(0).properties.unified_addressing, reason="requires unified addressing")
-def test_instantiate_and_execute_kernel_device_launch(sample_graphdef):
+def test_instantiate_and_execute_kernel_device_launch(init_cuda):
     """Kernel-only graph can be instantiated with device_launch flag."""
+    sample_graphdef = GraphDefinition()
     mod = compile_common_kernels()
     kernel = mod.get_kernel("empty_kernel")
     config = LaunchConfig(grid=1, block=1)
@@ -1045,8 +1044,9 @@ def test_instantiate_and_execute_kernel_device_launch(sample_graphdef):
 
 
 @pytest.mark.parametrize("inst_kwargs", _EXECUTE_OPTIONS)
-def test_instantiate_and_execute_kernel(sample_graphdef, inst_kwargs):
+def test_instantiate_and_execute_kernel(init_cuda, inst_kwargs):
     """Graph with kernel can be instantiated and executed."""
+    sample_graphdef = GraphDefinition()
     mod = compile_common_kernels()
     kernel = mod.get_kernel("empty_kernel")
     config = LaunchConfig(grid=1, block=1)
@@ -1059,8 +1059,9 @@ def test_instantiate_and_execute_kernel(sample_graphdef, inst_kwargs):
 
 
 @pytest.mark.parametrize("inst_kwargs", _EXECUTE_OPTIONS)
-def test_instantiate_and_execute_alloc_free(sample_graphdef, inst_kwargs):
+def test_instantiate_and_execute_alloc_free(init_cuda, inst_kwargs):
     """Graph with alloc/free can be executed."""
+    sample_graphdef = GraphDefinition()
     _skip_if_no_mempool()
     with xfail_on_graph_mempool_oom():
         alloc = sample_graphdef.allocate(ALLOC_SIZE)
@@ -1073,8 +1074,9 @@ def test_instantiate_and_execute_alloc_free(sample_graphdef, inst_kwargs):
 
 
 @pytest.mark.parametrize("inst_kwargs", _EXECUTE_OPTIONS)
-def test_instantiate_and_execute_memset(sample_graphdef, inst_kwargs):
+def test_instantiate_and_execute_memset(init_cuda, inst_kwargs):
     """Graph with alloc/memset/free can be executed."""
+    sample_graphdef = GraphDefinition()
     _skip_if_no_mempool()
     with xfail_on_graph_mempool_oom():
         alloc = sample_graphdef.allocate(ALLOC_SIZE)
@@ -1088,8 +1090,9 @@ def test_instantiate_and_execute_memset(sample_graphdef, inst_kwargs):
 
 
 @pytest.mark.parametrize("inst_kwargs", _EXECUTE_OPTIONS)
-def test_instantiate_and_execute_memcpy(sample_graphdef, inst_kwargs):
+def test_instantiate_and_execute_memcpy(init_cuda, inst_kwargs):
     """Graph with alloc/memset/memcpy/free can be executed and data is copied."""
+    sample_graphdef = GraphDefinition()
     _skip_if_no_mempool()
     import ctypes
 
@@ -1113,8 +1116,9 @@ def test_instantiate_and_execute_memcpy(sample_graphdef, inst_kwargs):
     assert all(b == 0xAB for b in host_buf)
 
 
-def test_instantiate_and_execute_child_graph(sample_graphdef):
+def test_instantiate_and_execute_child_graph(init_cuda):
     """Graph with embedded child graph can be executed."""
+    sample_graphdef = GraphDefinition()
     child = GraphDefinition()
     mod = compile_common_kernels()
     kernel = mod.get_kernel("empty_kernel")
@@ -1130,8 +1134,9 @@ def test_instantiate_and_execute_child_graph(sample_graphdef):
     stream.sync()
 
 
-def test_instantiate_and_execute_host_callback(sample_graphdef):
+def test_instantiate_and_execute_host_callback(init_cuda):
     """Graph with host callback can be executed and callback is invoked."""
+    sample_graphdef = GraphDefinition()
     results = []
 
     def my_callback():
@@ -1148,8 +1153,9 @@ def my_callback():
     assert results == [42]
 
 
-def test_instantiate_and_execute_host_callback_cfunc(sample_graphdef):
+def test_instantiate_and_execute_host_callback_cfunc(init_cuda):
     """Graph with ctypes function pointer callback can be executed."""
+    sample_graphdef = GraphDefinition()
     import ctypes
 
     CALLBACK = ctypes.CFUNCTYPE(None, ctypes.c_void_p)
@@ -1170,8 +1176,9 @@ def raw_fn(data):
     assert called[0]
 
 
-def test_host_callback_cfunc_with_user_data(sample_graphdef):
+def test_host_callback_cfunc_with_user_data(init_cuda):
     """Host callback with bytes user_data passes data to C function."""
+    sample_graphdef = GraphDefinition()
     import ctypes
 
     CALLBACK = ctypes.CFUNCTYPE(None, ctypes.c_void_p)
@@ -1192,8 +1199,9 @@ def read_byte(data):
     assert result[0] == 0xAB
 
 
-def test_host_callback_user_data_rejected_for_python_callable(sample_graphdef):
+def test_host_callback_user_data_rejected_for_python_callable(init_cuda):
     """user_data is rejected for Python callables."""
+    sample_graphdef = GraphDefinition()
     with pytest.raises(ValueError, match="user_data is only supported"):
         sample_graphdef.callback(lambda: None, user_data=b"hello")
 
@@ -1216,15 +1224,17 @@ def test_host_callback_user_data_rejected_for_python_callable(sample_graphdef):
 
 @pytest.mark.agent_authored(model="cursor-grok-4.5")
 @pytest.mark.parametrize("callback_type", _INCOMPATIBLE_CTYPES_HOST_CALLBACKS)
-def test_host_callback_ctypes_rejects_incompatible_signature(sample_graphdef, callback_type):
+def test_host_callback_ctypes_rejects_incompatible_signature(init_cuda, callback_type):
     """Incompatible ctypes prototypes are rejected before CUDA sees them."""
+    sample_graphdef = GraphDefinition()
     with pytest.raises(TypeError, match="CUhostFn"):
         sample_graphdef.callback(callback_type(0))
 
 
 @pytest.mark.agent_authored(model="cursor-grok-4.5")
-def test_host_callback_ctypes_update_rejects_incompatible_signature(sample_graphdef):
+def test_host_callback_ctypes_update_rejects_incompatible_signature(init_cuda):
     """HostCallbackNode.update applies the same ctypes ABI check."""
+    sample_graphdef = GraphDefinition()
     good_type = ctypes.CFUNCTYPE(None, ctypes.c_void_p)
     bad_type = ctypes.CFUNCTYPE(ctypes.c_int, ctypes.c_void_p)
 
@@ -1239,8 +1249,9 @@ def good(data):
 
 @pytest.mark.agent_authored(model="claude-opus-5")
 @pytest.mark.parametrize("callback_type", _COMPATIBLE_CTYPES_HOST_CALLBACKS)
-def test_host_callback_ctypes_accepts_equivalent_prototypes(sample_graphdef, callback_type):
+def test_host_callback_ctypes_accepts_equivalent_prototypes(init_cuda, callback_type):
     """Prototypes that declare CUhostFn are accepted and run."""
+    sample_graphdef = GraphDefinition()
     called = [False]
 
     @callback_type
@@ -1260,8 +1271,9 @@ def raw_fn(data):
 
 @pytest.mark.agent_authored(model="cursor-grok-4.5")
 @pytest.mark.skipif(sys.platform != "win32", reason="WINFUNCTYPE is Windows-only")
-def test_host_callback_ctypes_accepts_winfunctype(sample_graphdef):
+def test_host_callback_ctypes_accepts_winfunctype(init_cuda):
     """On Windows, WINFUNCTYPE matches CUDA_CB (__stdcall) and is accepted."""
+    sample_graphdef = GraphDefinition()
     callback_type = ctypes.WINFUNCTYPE(None, ctypes.c_void_p)
     called = [False]
 
@@ -1280,8 +1292,9 @@ def raw_fn(data):
     assert called[0]
 
 
-def test_instantiate_and_execute_event_record_wait(sample_graphdef):
+def test_instantiate_and_execute_event_record_wait(init_cuda):
     """Graph with event record and wait nodes can be executed."""
+    sample_graphdef = GraphDefinition()
     event = Device().create_event()
     rec = sample_graphdef.record(event)
     rec.wait(event)
@@ -1303,8 +1316,9 @@ def _skip_unless_cc_90():
         pytest.skip("Conditional node execution requires CC >= 9.0 (Hopper)")
 
 
-def test_instantiate_and_execute_if_then(sample_graphdef):
+def test_instantiate_and_execute_if_then(init_cuda):
     """If-conditional node: body executes only when condition is non-zero."""
+    sample_graphdef = GraphDefinition()
     _skip_unless_cc_90()
     _skip_if_no_mempool()
     import ctypes
@@ -1336,8 +1350,9 @@ def test_instantiate_and_execute_if_then(sample_graphdef):
     assert result[0] == 1
 
 
-def test_instantiate_and_execute_if_else(sample_graphdef):
+def test_instantiate_and_execute_if_else(init_cuda):
     """If-else node: then or else branch executes based on condition."""
+    sample_graphdef = GraphDefinition()
     _skip_unless_cc_90()
     _skip_if_no_mempool()
     import ctypes
@@ -1371,8 +1386,9 @@ def test_instantiate_and_execute_if_else(sample_graphdef):
     assert result[0] == 2
 
 
-def test_instantiate_and_execute_switch(sample_graphdef):
+def test_instantiate_and_execute_switch(init_cuda):
     """Switch node: selected branch executes based on condition value."""
+    sample_graphdef = GraphDefinition()
     _skip_unless_cc_90()
     _skip_if_no_mempool()
     import ctypes
@@ -1405,8 +1421,9 @@ def test_instantiate_and_execute_switch(sample_graphdef):
     assert result[0] == 1
 
 
-def test_conditional_node_type_preserved_by_nodes(sample_graphdef):
+def test_conditional_node_type_preserved_by_nodes(init_cuda):
     """Conditional nodes appear as ConditionalNode base when read back from graph."""
+    sample_graphdef = GraphDefinition()
     condition = try_create_condition(sample_graphdef)
     if_node = sample_graphdef.if_then(condition)
     assert isinstance(if_node, IfNode)
@@ -1422,8 +1439,10 @@ def test_conditional_node_type_preserved_by_nodes(sample_graphdef):
 # =============================================================================
 
 
-def test_debug_dot_print_creates_file(sample_graphdef, dot_file):
+def test_debug_dot_print_creates_file(init_cuda, tmp_path):
     """debug_dot_print writes a DOT file."""
+    sample_graphdef = GraphDefinition()
+    dot_file = tmp_path / "graph.dot"
     _skip_if_no_mempool()
     with xfail_on_graph_mempool_oom():
         sample_graphdef.allocate(ALLOC_SIZE)
@@ -1433,8 +1452,10 @@ def test_debug_dot_print_creates_file(sample_graphdef, dot_file):
     assert "digraph" in content
 
 
-def test_debug_dot_print_with_options(sample_graphdef, dot_file):
+def test_debug_dot_print_with_options(init_cuda, tmp_path):
     """debug_dot_print accepts GraphDebugPrintOptions."""
+    sample_graphdef = GraphDefinition()
+    dot_file = tmp_path / "graph.dot"
     _skip_if_no_mempool()
     with xfail_on_graph_mempool_oom():
         sample_graphdef.allocate(ALLOC_SIZE)
@@ -1443,8 +1464,10 @@ def test_debug_dot_print_with_options(sample_graphdef, dot_file):
     assert dot_file.exists()
 
 
-def test_debug_dot_print_invalid_options(sample_graphdef, dot_file):
+def test_debug_dot_print_invalid_options(init_cuda, tmp_path):
     """debug_dot_print rejects invalid options type."""
+    sample_graphdef = GraphDefinition()
+    dot_file = tmp_path / "graph.dot"
     _skip_if_no_mempool()
     with xfail_on_graph_mempool_oom():
         sample_graphdef.allocate(ALLOC_SIZE)
diff --git a/cuda_core/tests/graph/test_graph_definition_lifetime.py b/cuda_core/tests/graph/test_graph_definition_lifetime.py
index c9ff492a2a8..fc82b7a04eb 100644
--- a/cuda_core/tests/graph/test_graph_definition_lifetime.py
+++ b/cuda_core/tests/graph/test_graph_definition_lifetime.py
@@ -685,6 +685,7 @@ def test_event_survives_graph_clone_and_execution(init_cuda):
 # =============================================================================
 
 
+@pytest.mark.thread_unsafe(reason="asserts cleanup on main thread")
 @pytest.mark.agent_authored(model="gpt-5.6")
 def test_user_object_cleanup_is_coalesced_on_python_thread(init_cuda):
     """More than 32 CUDA callbacks drain through one main-thread pending call."""
@@ -1412,6 +1413,7 @@ def test_memcpy_buffer_survives_close(init_cuda):
     assert list(out) == [0xCD] * 4
 
 
+@pytest.mark.thread_unsafe(reason="deferred cleanup on main thread which would wait")
 @pytest.mark.agent_authored(model="claude-opus-4.8")
 def test_memcpy_buffer_allocations_released_after_graph_destroyed(init_cuda):
     """Destroying the graph frees both memcpy operand allocations.
diff --git a/cuda_core/tests/graph/test_graph_memory_resource.py b/cuda_core/tests/graph/test_graph_memory_resource.py
index 7b9c2aaf883..22757ca7556 100644
--- a/cuda_core/tests/graph/test_graph_memory_resource.py
+++ b/cuda_core/tests/graph/test_graph_memory_resource.py
@@ -5,7 +5,7 @@
 """Tests for GraphMemoryResource allocation and attributes during graph capture."""
 
 import pytest
-from helpers.buffers import compare_buffer_to_constant, make_scratch_buffer, set_buffer
+from helpers.buffers import compare_buffer_to_constant, make_scratch_buffer, set_buffer, thread_unsafe_on_windows
 from helpers.memory import xfail_on_graph_mempool_oom
 
 from cuda.core import (
@@ -21,6 +21,13 @@
 from cuda.core.graph import GraphCompleteOptions
 from cuda_python_test_helpers import IS_WINDOWS, IS_WSL
 
+# NOTE(seberg): "global" mode seems thread-unsafe even when working on stream
+_GRAPH_MODES = [
+    pytest.param("global", marks=pytest.mark.thread_unsafe(reason="gb instances share stream unsafely")),
+    "thread_local",
+    "relaxed",
+]
+
 
 def _common_kernels_alloc():
     code = """
@@ -80,8 +87,9 @@ def free(self, buffers):
         self.stream.sync()
 
 
-@pytest.mark.parametrize("mode", ["no_graph", "global", "thread_local", "relaxed"])
+@pytest.mark.parametrize("mode", ["no_graph"] + _GRAPH_MODES)
 @pytest.mark.parametrize("action", ["incr", "fill"])
+@thread_unsafe_on_windows
 def test_graph_alloc(mempool_device, mode, action):
     """Test basic graph capture with memory allocated and deallocated by
     GraphMemoryResource.
@@ -130,7 +138,7 @@ def apply_kernels(mr, stream, out):
         assert compare_buffer_to_constant(out, 3)
     else:
         # Capture work, then upload and launch.
-        gb = device.create_graph_builder().begin_building(mode)
+        gb = stream.create_graph_builder().begin_building(mode)
         with xfail_on_graph_mempool_oom(device):
             apply_kernels(mr=gmr, stream=gb, out=out)
             graph = gb.end_building().complete()
@@ -150,7 +158,8 @@ def apply_kernels(mr, stream, out):
 
 
 @pytest.mark.skipif(IS_WINDOWS or IS_WSL, reason="auto_free_on_launch not supported on Windows")
-@pytest.mark.parametrize("mode", ["global", "thread_local", "relaxed"])
+@pytest.mark.parametrize("mode", _GRAPH_MODES)
+@thread_unsafe_on_windows
 def test_graph_alloc_with_output(mempool_device, mode):
     """Test for memory allocated in a graph being used outside the graph."""
     NBYTES = 64
@@ -168,7 +177,7 @@ def test_graph_alloc_with_output(mempool_device, mode):
     # Construct a graph to copy and increment the input. It returns a new
     # buffer allocated within the graph.  The auto_free_on_launch option
     # is required to properly use the output buffer.
-    gb = device.create_graph_builder().begin_building(mode)
+    gb = stream.create_graph_builder().begin_building(mode)
     with xfail_on_graph_mempool_oom(device):
         out = gmr.allocate(NBYTES, stream=gb)
         out.copy_from(in_, stream=gb)
@@ -195,7 +204,8 @@ def test_graph_alloc_with_output(mempool_device, mode):
     assert compare_buffer_to_constant(out, 6)
 
 
-@pytest.mark.parametrize("mode", ["global", "thread_local", "relaxed"])
+@pytest.mark.parametrize("mode", _GRAPH_MODES)
+@pytest.mark.thread_unsafe(reason="gb instances share default stream")
 def test_graph_mem_alloc_zero(mempool_device, mode):
     device = mempool_device
     gb = device.create_graph_builder().begin_building(mode)
@@ -213,7 +223,8 @@ def test_graph_mem_alloc_zero(mempool_device, mode):
     assert buffer.device_id == int(device)
 
 
-@pytest.mark.parametrize("mode", ["global", "thread_local", "relaxed"])
+@pytest.mark.parametrize("mode", _GRAPH_MODES)
+@pytest.mark.thread_unsafe(reason="GMR is shared, so high mark is global")
 def test_graph_mem_set_attributes(mempool_device, mode):
     device = mempool_device
     stream = device.create_stream()
@@ -265,7 +276,7 @@ def test_graph_mem_set_attributes(mempool_device, mode):
     mman.reset()
 
 
-@pytest.mark.parametrize("mode", ["global", "thread_local", "relaxed"])
+@pytest.mark.parametrize("mode", _GRAPH_MODES)
 def test_gmr_check_capture_state(mempool_device, mode):
     """
     Test expected errors (and non-errors) using GraphMemoryResource with graph
@@ -284,7 +295,7 @@ def test_gmr_check_capture_state(mempool_device, mode):
         gmr.allocate(1, stream=stream)
 
     # Capturing
-    gb = device.create_graph_builder().begin_building(mode=mode)
+    gb = stream.create_graph_builder().begin_building(mode=mode)
     with xfail_on_graph_mempool_oom(device):
         gmr.allocate(1, stream=gb)  # no error
         gb.end_building().complete()
@@ -320,7 +331,7 @@ def test_graph_memory_resource_attributes_repr(mempool_device):
     assert "used_mem_high=" in r
 
 
-@pytest.mark.parametrize("mode", ["global", "thread_local", "relaxed"])
+@pytest.mark.parametrize("mode", _GRAPH_MODES)
 def test_dmr_check_capture_state(mempool_device, mode):
     """
     Test expected errors (and non-errors) using DeviceMemoryResource with graph
@@ -334,7 +345,7 @@ def test_dmr_check_capture_state(mempool_device, mode):
     dmr.allocate(1, stream=stream).close()  # no error
 
     # Capturing
-    gb = device.create_graph_builder().begin_building(mode=mode)
+    gb = stream.create_graph_builder().begin_building(mode=mode)
     with pytest.raises(
         RuntimeError,
         match=r"cannot perform memory operations on a capturing "
diff --git a/cuda_core/tests/graph/test_graph_update.py b/cuda_core/tests/graph/test_graph_update.py
index 54a04863cf4..c95e331913a 100644
--- a/cuda_core/tests/graph/test_graph_update.py
+++ b/cuda_core/tests/graph/test_graph_update.py
@@ -196,6 +196,7 @@ def new_callback():
     assert called == ["new"]
 
 
+@pytest.mark.thread_unsafe(reason="deferred cleanup on main thread which would wait")
 @pytest.mark.agent_authored(model="gpt-5.6")
 def test_failed_graph_update_does_not_adopt_attachments(init_cuda):
     """A rejected source graph keeps ownership separate from the exec."""
diff --git a/cuda_core/tests/helpers/buffers.py b/cuda_core/tests/helpers/buffers.py
index aefb4f04193..2e71a2a199a 100644
--- a/cuda_core/tests/helpers/buffers.py
+++ b/cuda_core/tests/helpers/buffers.py
@@ -3,11 +3,13 @@
 
 import ctypes
 
+import pytest
+
 from cuda.core import Buffer, Device, MemoryResource
 from cuda.core._stream import Stream_accept
 from cuda.core._utils.cuda_utils import driver, handle_return
 
-from . import libc
+from . import IS_WINDOWS, IS_WSL, libc
 
 __all__ = [
     "DummyDeviceMemoryResource",
@@ -18,9 +20,19 @@
     "compare_equal_buffers",
     "make_instrumented_memory_resource",
     "make_scratch_buffer",
+    "thread_unsafe_on_windows",
 ]
 
 
+def thread_unsafe_on_windows(func):
+    # Tests that use these buffers and access the memory on the host are
+    # thread-unsafe on windows. On windows the GPU must be fully quiescent for host
+    # access to be safe and with threaded tests that would require a barrier.
+    if IS_WINDOWS or IS_WSL:
+        return pytest.mark.thread_unsafe(reason="windows host-access unsafe while GPU is working")
+    return func
+
+
 class StubMemoryResource(MemoryResource):
     """Device-only memory resource for tests that supply a fake pointer."""
 
diff --git a/cuda_core/tests/helpers/latch.py b/cuda_core/tests/helpers/latch.py
index c28fb222641..7702c5f617e 100644
--- a/cuda_core/tests/helpers/latch.py
+++ b/cuda_core/tests/helpers/latch.py
@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 import ctypes
diff --git a/cuda_core/tests/memory/test_managed_ops.py b/cuda_core/tests/memory/test_managed_ops.py
index cb686026c7e..99053e1f8c9 100644
--- a/cuda_core/tests/memory/test_managed_ops.py
+++ b/cuda_core/tests/memory/test_managed_ops.py
@@ -365,6 +365,7 @@ def test_from_handle(self, init_cuda):
         finally:
             plain.close()
 
+    @pytest.mark.thread_unsafe(reason="external_managed_buffer is shared between threads")
     def test_read_mostly_roundtrip(self, external_managed_buffer):
         buf = external_managed_buffer
         assert buf.read_mostly is False
@@ -373,6 +374,7 @@ def test_read_mostly_roundtrip(self, external_managed_buffer):
         buf.read_mostly = False
         assert buf.read_mostly is False
 
+    @pytest.mark.thread_unsafe(reason="external_managed_buffer is shared between threads")
     def test_preferred_location_roundtrip(self, location_ops_device, external_managed_buffer):
         device = location_ops_device
         buf = external_managed_buffer
@@ -387,6 +389,7 @@ def test_preferred_location_roundtrip(self, location_ops_device, external_manage
         buf.preferred_location = None
         assert buf.preferred_location is None
 
+    @pytest.mark.thread_unsafe(reason="external_managed_buffer is shared between threads")
     def test_preferred_location_roundtrip_host_numa(self, location_ops_device):
         """Host(numa_id=N) round-trips correctly on CUDA 13 builds."""
         from cuda.core._utils.version import binding_version
@@ -407,6 +410,7 @@ def test_preferred_location_roundtrip_host_numa(self, location_ops_device):
         finally:
             plain.close()
 
+    @pytest.mark.thread_unsafe(reason="external_managed_buffer is shared between threads")
     def test_accessed_by_add_discard(self, location_ops_device, external_managed_buffer):
         device = location_ops_device
         buf = external_managed_buffer
@@ -418,6 +422,7 @@ def test_accessed_by_add_discard(self, location_ops_device, external_managed_buf
         buf.accessed_by.discard(device)
         assert device not in buf.accessed_by
 
+    @pytest.mark.thread_unsafe(reason="external_managed_buffer is shared between threads")
     def test_accessed_by_mutable_set_interface(self, location_ops_device, external_managed_buffer):
         """Full MutableSet conformance pass on AccessedBySetProxy.
 
@@ -437,6 +442,7 @@ def test_accessed_by_mutable_set_interface(self, location_ops_device, external_m
             non_member=Host(numa_id=0),
         )
 
+    @pytest.mark.thread_unsafe(reason="external_managed_buffer is shared between threads")
     def test_accessed_by_set_assignment(self, location_ops_device, external_managed_buffer):
         device = location_ops_device
         buf = external_managed_buffer
diff --git a/cuda_core/tests/system/test_system_device.py b/cuda_core/tests/system/test_system_device.py
index 028b7e7adf7..078ca5a835c 100644
--- a/cuda_core/tests/system/test_system_device.py
+++ b/cuda_core/tests/system/test_system_device.py
@@ -377,6 +377,7 @@ def test_c2c_mode_enabled(subtests):
 
 
 @pytest.mark.skipif(helpers.IS_WSL or helpers.IS_WINDOWS, reason="Persistence mode not supported on WSL or Windows")
+@pytest.mark.thread_unsafe(reason="device persistence mode is global state")
 def test_persistence_mode_enabled(subtests):
     for device in system.Device.get_all_devices():
         with subtests.test(device_index=device.index):
diff --git a/cuda_core/tests/test_event.py b/cuda_core/tests/test_event.py
index 3e765a9e0d0..bbe5a16efac 100644
--- a/cuda_core/tests/test_event.py
+++ b/cuda_core/tests/test_event.py
@@ -119,6 +119,7 @@ def test_error_timing_recorded():
 
 
 @pytest.mark.skipif(Device().compute_capability.major < 7, reason="__nanosleep is only available starting Volta (sm70)")
+@pytest.mark.thread_unsafe(reason="requires a barrier wait to avoid overlapping pinned latch allocations")
 def test_error_timing_incomplete():
     device = Device()
     device.set_current()
@@ -223,6 +224,7 @@ def test_event_ipc_descriptor_non_ipc(init_cuda):
 
 
 @pytest.mark.skipif(Device().compute_capability.major < 7, reason="__nanosleep is only available starting Volta (sm70)")
+@pytest.mark.thread_unsafe(reason="requires a barrier wait to avoid overlapping pinned latch allocations")
 def test_event_is_done_false(init_cuda):
     """Event.is_done returns False when captured work has not yet completed."""
     device = Device()
diff --git a/cuda_core/tests/test_helpers.py b/cuda_core/tests/test_helpers.py
index ed1e3bc6e74..8c55b924d77 100644
--- a/cuda_core/tests/test_helpers.py
+++ b/cuda_core/tests/test_helpers.py
@@ -6,7 +6,7 @@
 import types
 
 import pytest
-from helpers.buffers import PatternGen, compare_equal_buffers, make_scratch_buffer
+from helpers.buffers import PatternGen, compare_equal_buffers, make_scratch_buffer, thread_unsafe_on_windows
 from helpers.latch import LatchKernel
 from helpers.logging import TimestampedLogger
 from helpers.oom_diagnostics import (
@@ -29,6 +29,7 @@
 
 
 @pytest.mark.skipif(Device().compute_capability.major < 7, reason="__nanosleep is only available starting Volta (sm70)")
+@pytest.mark.thread_unsafe(reason="requires a barrier wait to avoid overlapping pinned latch allocations")
 def test_latchkernel():
     """Test LatchKernel."""
     log = TimestampedLogger(enabled=ENABLE_LOGGING)
@@ -63,6 +64,7 @@ def test_latchkernel():
     under_compute_sanitizer(),
     reason="Too slow under compute-sanitizer (UVM-heavy test).",
 )
+@thread_unsafe_on_windows
 def test_patterngen_seeds():
     """Test PatternGen with seed argument."""
     device = Device()
@@ -81,6 +83,7 @@ def test_patterngen_seeds():
                 pgen.verify_buffer(buffer, seed=j)
 
 
+@thread_unsafe_on_windows
 def test_patterngen_values():
     """Test PatternGen with value argument, also compare_equal_buffers."""
     device = Device()
diff --git a/cuda_core/tests/test_memory.py b/cuda_core/tests/test_memory.py
index 99aaf938721..21f1ecab5ea 100644
--- a/cuda_core/tests/test_memory.py
+++ b/cuda_core/tests/test_memory.py
@@ -20,6 +20,7 @@
     DummyUnifiedMemoryResource,
     StubMemoryResource,
     make_instrumented_memory_resource,
+    thread_unsafe_on_windows,
 )
 from helpers.constants import POOL_SIZE
 from helpers.memory import (
@@ -283,9 +284,8 @@ def _pattern_bytes(value) -> bytes:
 
 
 @pytest.fixture(params=["device", "unified", "pinned"])
-def fill_env(request):
-    device = Device()
-    device.set_current()
+def fill_env(request, init_cuda):
+    device = init_cuda
     if request.param == "device":
         mr = DummyDeviceMemoryResource(device)
     elif request.param == "unified":
@@ -348,6 +348,7 @@ def fill_env(request):
     )
 
 
+@thread_unsafe_on_windows
 @pytest.mark.parametrize("value,size,exc", _FILL_CASES)
 def test_buffer_fill(fill_env, value, size, exc):
     device, mr = fill_env
@@ -1519,6 +1520,8 @@ def test_managed_memory_resource_with_options(init_cuda):
     device.sync()
     dst_buffer.close()
     src_buffer.close()
+    # TODO(seberg): 2026-06: mr close may be unsafe with incomplete `buf.close()`
+    device.sync()
 
 
 def test_managed_memory_resource_preferred_location_default(init_cuda):
diff --git a/cuda_core/tests/test_multiprocessing_warning.py b/cuda_core/tests/test_multiprocessing_warning.py
index 1ddb53edb0f..2eb4aa55de9 100644
--- a/cuda_core/tests/test_multiprocessing_warning.py
+++ b/cuda_core/tests/test_multiprocessing_warning.py
@@ -12,6 +12,7 @@
 import warnings
 from unittest.mock import patch
 
+import pytest
 from helpers.constants import POOL_SIZE
 
 from cuda.core import DeviceMemoryResource, DeviceMemoryResourceOptions, EventOptions
@@ -20,6 +21,9 @@
 from cuda.core._memory._ipc import _reduce_allocation_handle
 from cuda.core._utils.cuda_utils import check_multiprocessing_start_method, reset_fork_warning
 
+# We could move these to a (session) fixtures
+pytestmark = pytest.mark.thread_unsafe(reason="all tests use unittest.mock.patch")
+
 
 def test_warn_on_fork_method_device_memory_resource(ipc_device):
     """Test that warning is emitted when DeviceMemoryResource is pickled with fork method."""
diff --git a/cuda_core/tests/test_object_protocols.py b/cuda_core/tests/test_object_protocols.py
index cd664501d06..b86cf5cf7e3 100644
--- a/cuda_core/tests/test_object_protocols.py
+++ b/cuda_core/tests/test_object_protocols.py
@@ -526,6 +526,25 @@ def sample_switch_node_alt(sample_graphdef):
     return sample_graphdef.switch(condition, 3)
 
 
+# Resolve fixture names during pytest setup, before pytest-run-parallel starts
+# worker threads. The workers then share the resolved, read-only test object.
+
+
+@pytest.fixture
+def sample_object(request):
+    return request.getfixturevalue(request.param)
+
+
+@pytest.fixture
+def sample_object_a(request):
+    return request.getfixturevalue(request.param)
+
+
+@pytest.fixture
+def sample_object_b(request):
+    return request.getfixturevalue(request.param)
+
+
 # =============================================================================
 # Type groupings
 # =============================================================================
@@ -722,12 +741,11 @@ def sample_switch_node_alt(sample_graphdef):
 # =============================================================================
 
 
-@pytest.mark.parametrize("fixture_name", WEAKREF_TYPES)
-def test_weakref_supported(fixture_name, request):
+@pytest.mark.parametrize("sample_object", WEAKREF_TYPES, indirect=True)
+def test_weakref_supported(sample_object):
     """Object supports weak references."""
-    obj = request.getfixturevalue(fixture_name)
-    ref = weakref.ref(obj)
-    assert ref() is obj
+    ref = weakref.ref(sample_object)
+    assert ref() is sample_object
 
 
 # =============================================================================
@@ -735,27 +753,22 @@ def test_weakref_supported(fixture_name, request):
 # =============================================================================
 
 
-@pytest.mark.parametrize("fixture_name", HASH_TYPES)
-def test_hash_consistency(fixture_name, request):
+@pytest.mark.parametrize("sample_object", HASH_TYPES, indirect=True)
+def test_hash_consistency(sample_object):
     """Hash is consistent across multiple calls."""
-    obj = request.getfixturevalue(fixture_name)
-    assert hash(obj) == hash(obj)
+    assert hash(sample_object) == hash(sample_object)
 
 
-@pytest.mark.parametrize("a_name,b_name", SAME_TYPE_PAIRS)
-def test_hash_distinct_same_type(a_name, b_name, request):
+@pytest.mark.parametrize("sample_object_a,sample_object_b", SAME_TYPE_PAIRS, indirect=True)
+def test_hash_distinct_same_type(sample_object_a, sample_object_b):
     """Distinct objects of the same type have different hashes."""
-    obj_a = request.getfixturevalue(a_name)
-    obj_b = request.getfixturevalue(b_name)
-    assert hash(obj_a) != hash(obj_b)  # extremely unlikely
+    assert hash(sample_object_a) != hash(sample_object_b)  # extremely unlikely
 
 
-@pytest.mark.parametrize("a_name,b_name", list(itertools.combinations(HASH_TYPES, 2)))
-def test_hash_distinct_cross_type(a_name, b_name, request):
+@pytest.mark.parametrize("sample_object_a,sample_object_b", itertools.combinations(HASH_TYPES, 2), indirect=True)
+def test_hash_distinct_cross_type(sample_object_a, sample_object_b):
     """Distinct objects of different types have different hashes."""
-    obj_a = request.getfixturevalue(a_name)
-    obj_b = request.getfixturevalue(b_name)
-    assert hash(obj_a) != hash(obj_b)  # extremely unlikely
+    assert hash(sample_object_a) != hash(sample_object_b)  # extremely unlikely
 
 
 # =============================================================================
@@ -763,41 +776,35 @@ def test_hash_distinct_cross_type(a_name, b_name, request):
 # =============================================================================
 
 
-@pytest.mark.parametrize("fixture_name", EQ_TYPES)
-def test_equality_basic(fixture_name, request):
+@pytest.mark.parametrize("sample_object", EQ_TYPES, indirect=True)
+def test_equality_basic(sample_object):
     """Object equality: reflexive, not equal to None or other types."""
-    obj = request.getfixturevalue(fixture_name)
-    assert obj == obj
-    assert obj is not None
-    assert obj != "string"
-    if hasattr(obj, "handle"):
-        assert obj != obj.handle
+    assert sample_object == sample_object
+    assert sample_object is not None
+    assert sample_object != "string"
+    if hasattr(sample_object, "handle"):
+        assert sample_object != sample_object.handle
 
 
-@pytest.mark.parametrize("a_name,b_name", list(itertools.combinations(EQ_TYPES, 2)))
-def test_no_cross_type_equality(a_name, b_name, request):
+@pytest.mark.parametrize("sample_object_a,sample_object_b", itertools.combinations(EQ_TYPES, 2), indirect=True)
+def test_no_cross_type_equality(sample_object_a, sample_object_b):
     """No two distinct objects of different types should compare equal."""
-    obj_a = request.getfixturevalue(a_name)
-    obj_b = request.getfixturevalue(b_name)
-    assert obj_a != obj_b
+    assert sample_object_a != sample_object_b
 
 
-@pytest.mark.parametrize("a_name,b_name", SAME_TYPE_PAIRS)
-def test_same_type_inequality(a_name, b_name, request):
+@pytest.mark.parametrize("sample_object_a,sample_object_b", SAME_TYPE_PAIRS, indirect=True)
+def test_same_type_inequality(sample_object_a, sample_object_b):
     """Two distinct objects of the same type should not compare equal."""
-    obj_a = request.getfixturevalue(a_name)
-    obj_b = request.getfixturevalue(b_name)
-    assert obj_a is not obj_b
-    assert obj_a != obj_b
+    assert sample_object_a is not sample_object_b
+    assert sample_object_a != sample_object_b
 
 
-@pytest.mark.parametrize("fixture_name,copy_fn", FROM_HANDLE_COPIES)
-def test_equality_same_handle(fixture_name, copy_fn, request):
+@pytest.mark.parametrize("sample_object,copy_fn", FROM_HANDLE_COPIES, indirect=["sample_object"])
+def test_equality_same_handle(sample_object, copy_fn):
     """Two wrappers around the same handle should compare equal."""
-    obj = request.getfixturevalue(fixture_name)
-    obj2 = copy_fn(obj)
-    assert obj == obj2
-    assert hash(obj) == hash(obj2)
+    obj2 = copy_fn(sample_object)
+    assert sample_object == obj2
+    assert hash(sample_object) == hash(obj2)
 
 
 # =============================================================================
@@ -805,48 +812,43 @@ def test_equality_same_handle(fixture_name, copy_fn, request):
 # =============================================================================
 
 
-@pytest.mark.parametrize("fixture_name", DICT_KEY_TYPES)
-def test_usable_as_dict_key(fixture_name, request):
+@pytest.mark.parametrize("sample_object", DICT_KEY_TYPES, indirect=True)
+def test_usable_as_dict_key(sample_object):
     """Object can be used as a dictionary key."""
-    obj = request.getfixturevalue(fixture_name)
-    d = {obj: "value"}
-    assert d[obj] == "value"
-    assert obj in d
+    d = {sample_object: "value"}
+    assert d[sample_object] == "value"
+    assert sample_object in d
 
 
-@pytest.mark.parametrize("fixture_name", DICT_KEY_TYPES)
-def test_usable_in_set(fixture_name, request):
+@pytest.mark.parametrize("sample_object", DICT_KEY_TYPES, indirect=True)
+def test_usable_in_set(sample_object):
     """Object can be added to a set."""
-    obj = request.getfixturevalue(fixture_name)
-    s = {obj}
-    assert obj in s
+    s = {sample_object}
+    assert sample_object in s
 
 
-@pytest.mark.parametrize("fixture_name", WEAKREF_TYPES)
-def test_usable_in_weak_value_dict(fixture_name, request):
+@pytest.mark.parametrize("sample_object", WEAKREF_TYPES, indirect=True)
+def test_usable_in_weak_value_dict(sample_object):
     """Object can be used as a WeakValueDictionary value."""
-    obj = request.getfixturevalue(fixture_name)
     wvd = weakref.WeakValueDictionary()
-    wvd["key"] = obj
-    assert wvd["key"] is obj
+    wvd["key"] = sample_object
+    assert wvd["key"] is sample_object
 
 
-@pytest.mark.parametrize("fixture_name", WEAK_KEY_TYPES)
-def test_usable_in_weak_key_dict(fixture_name, request):
+@pytest.mark.parametrize("sample_object", WEAK_KEY_TYPES, indirect=True)
+def test_usable_in_weak_key_dict(sample_object):
     """Object can be used as a WeakKeyDictionary key."""
-    obj = request.getfixturevalue(fixture_name)
     wkd = weakref.WeakKeyDictionary()
-    wkd[obj] = "value"
-    assert wkd[obj] == "value"
+    wkd[sample_object] = "value"
+    assert wkd[sample_object] == "value"
 
 
-@pytest.mark.parametrize("fixture_name", WEAK_KEY_TYPES)
-def test_usable_in_weak_set(fixture_name, request):
+@pytest.mark.parametrize("sample_object", WEAK_KEY_TYPES, indirect=True)
+def test_usable_in_weak_set(sample_object):
     """Object can be added to a WeakSet."""
-    obj = request.getfixturevalue(fixture_name)
     ws = weakref.WeakSet()
-    ws.add(obj)
-    assert obj in ws
+    ws.add(sample_object)
+    assert sample_object in ws
 
 
 # =============================================================================
@@ -854,12 +856,10 @@ def test_usable_in_weak_set(fixture_name, request):
 # =============================================================================
 
 
-@pytest.mark.parametrize("fixture_name,pattern", REPR_PATTERNS)
-def test_repr_format(fixture_name, pattern, request):
+@pytest.mark.parametrize("sample_object,pattern", REPR_PATTERNS, indirect=["sample_object"])
+def test_repr_format(sample_object, pattern):
     """repr() returns a properly formatted string."""
-    obj = request.getfixturevalue(fixture_name)
-    result = repr(obj)
-    assert re.fullmatch(pattern, result)
+    assert re.fullmatch(pattern, repr(sample_object))
 
 
 # =============================================================================
@@ -868,10 +868,9 @@ def test_repr_format(fixture_name, pattern, request):
 
 
 @pytest.mark.parametrize("pickle_module", PICKLE_MODULES)
-@pytest.mark.parametrize("fixture_name", PICKLE_TYPES)
-def test_pickle_roundtrip(fixture_name, pickle_module, request):
+@pytest.mark.parametrize("sample_object", PICKLE_TYPES, indirect=True)
+def test_pickle_roundtrip(sample_object, pickle_module):
     """Object survives a pickle/cloudpickle roundtrip."""
     mod = pytest.importorskip(pickle_module)
-    obj = request.getfixturevalue(fixture_name)
-    result = mod.loads(mod.dumps(obj))
-    assert type(result) is type(obj)
+    result = mod.loads(mod.dumps(sample_object))
+    assert type(result) is type(sample_object)