tests: Mark tests as unsafe or limit number of threads. (#2229)

* TST: Mark tests as thread-unsafe or limit the number of threads

- thread_unsafe: nvml init ref-count, graphMem attr, mock-based tests,
  OpenGL, peer-access pool state, multiprocessing warning, program-cache
  race reproduction, and functools.cache mutation tests
- parallel_threads_limit: IPC / worker-pool tests that spawn subprocesses
  or open file descriptors (limit 4), example tests (limit 8), and the
  event-registration test whose timeouts are slow

Signed-off-by: Sebastian Berg <sebastianb@nvidia.com>

* TST: Add two more thread-unsafe markers

* TST: Mark `mocker` as a thread-unsafe fixture

* Remove unsafe marker from test_multiprocessing_warning (can do the fixtures)

---------

Signed-off-by: Sebastian Berg <sebastianb@nvidia.com>

Author: seberg

cuda_bindings/tests/nvml/test_device.py

@@ -120,6 +120,7 @@ def test_read_prm_counters(all_devices):
         assert len(read_counters) == 5
 
 
+@pytest.mark.thread_unsafe(reason="API appears to be thread-unsafe (2026-06)")
 def test_read_write_prm(all_devices):
     for device in all_devices:
         # Docs say supported in BLACKWELL or later

cuda_bindings/tests/nvml/test_init.py

@@ -42,6 +42,7 @@ def get_architecture_name(arch):
 
 
 @pytest.mark.skipif(sys.platform == "win32", reason="Test not supported on Windows")
+@pytest.mark.thread_unsafe(reason="nvml init affects other threads")
 def test_init_ref_count():
     """
     Verifies that we can call NVML shutdown and init(2) multiple times, and that ref counting works

cuda_bindings/tests/test_cuda.py

@@ -456,6 +456,7 @@ def test_cuda_mem_range_attr(device):
 
 
 @pytest.mark.skipif(driverVersionLessThan(11040) or not supportsMemoryPool(), reason="Mempool for graphs not supported")
+@pytest.mark.thread_unsafe(reason="used high memory can be higher if threaded.")
 def test_cuda_graphMem_attr(device):
     err, stream = cuda.cuStreamCreate(0)
     assert err == cuda.CUresult.CUDA_SUCCESS

cuda_core/tests/example_tests/test_basic_examples.py

@@ -100,6 +100,7 @@ def has_recent_memory_pool_support() -> bool:
 
 
 @pytest.mark.parametrize("example", sample_files)
+@pytest.mark.parallel_threads_limit(8)
 def test_example(example):
     example_path = os.path.join(samples_path, example)
     has_package_requirements_or_skip(example_path)

cuda_core/tests/memory_ipc/test_errors.py

@@ -17,6 +17,10 @@
 POOL_SIZE = 2097152
 
 
+# these tests spawn new processes and files which fails for very many threads
+pytestmark = pytest.mark.parallel_threads_limit(4)
+
+
 def test_outer_timeout_marker_is_applied(request):
     """Verify that memory_ipc/conftest.py applies the outer pytest-timeout marker.
 

cuda_core/tests/memory_ipc/test_event_ipc.py

@@ -16,6 +16,10 @@
 NBYTES = 64
 
 
+# these tetss spawn new processes and files which fails for very many threads
+pytestmark = pytest.mark.parallel_threads_limit(4)
+
+
 @pytest.mark.skipif(Device().compute_capability.major < 7, reason="__nanosleep is only available starting Volta (sm70)")
 class TestEventIpc:
     """Check the basic usage of IPC-enabled events with a latch kernel."""

cuda_core/tests/memory_ipc/test_ipc_duplicate_import.py

@@ -24,6 +24,9 @@
 
 ENABLE_LOGGING = False  # Set True for test debugging and development
 
+# these tests spawn new processes and files which fails for very many threads
+pytestmark = pytest.mark.parallel_threads_limit(4)
+
 
 def child_main(log, queue):
     log.prefix = " child: "

cuda_core/tests/memory_ipc/test_leaks.py

@@ -23,6 +23,8 @@
     not USING_FDS or not HAVE_PSUTIL, reason="mempool allocation handle is not using fds or psutil is unavailable"
 )
 
+pytestmark = pytest.mark.thread_unsafe(reason="Tests number of fds which is shared.")
+
 
 @pytest.mark.flaky(reruns=2)
 @skip_if_unrunnable

cuda_core/tests/memory_ipc/test_memory_ipc.py

@@ -14,6 +14,9 @@
 NWORKERS = 2
 NTASKS = 2
 
+# these tests spawn new processes and files which fails for very many threads
+pytestmark = pytest.mark.parallel_threads_limit(4)
+
 
 class TestIpcMempool:
     @pytest.mark.flaky(reruns=2)

cuda_core/tests/memory_ipc/test_peer_access.py

@@ -14,6 +14,9 @@
 NBYTES = 64
 POOL_SIZE = 2097152
 
+# these tests spawn new processes and files which fails for very many threads
+pytestmark = pytest.mark.parallel_threads_limit(4)
+
 
 class TestPeerAccessNotPreservedOnImport:
     """