[cuda.compute]: Add pytest-run-parallel

ci/test_cuda_compute_minimal_python.sh

@@ -30,7 +30,7 @@ fi
 # full cu* extras because those pull in numba/numba-cuda.
 CUDA_CCCL_WHEEL_PATH="$(ls "${wheelhouse_dir}"/cuda_cccl-*.whl)"
 python -m pip install "${CUDA_CCCL_WHEEL_PATH}[minimal-cu${cuda_major_version}]"
-python -m pip install pytest pytest-xdist
+python -m pip install pytest pytest-xdist pytest-run-parallel
 
 cd "${repo_root}/python/cuda_cccl/tests/"
 python -m pytest -n 6 -v compute/test_no_numba.py
@@ -44,4 +44,24 @@ if [[ "${py_version}" == "3.14t" ]]; then
     compute/test_free_threading_stress.py \
     compute/test_multi_cc_serialization.py::test_aot_build_result_load_failure_is_shared_and_retryable \
     compute/test_multi_cc_serialization.py::test_aot_serialization_waits_for_canonical_first_load
+
+  # Broad thread-safety sweep (pytest-run-parallel): re-run the numba-free
+  # functional suite with each test executed concurrently across threads
+  # (barrier-synchronized start), stressing the process-wide build cache,
+  # single-flight coordination, and the Cython bindings from many threads at
+  # once. Complements test_free_threading_stress.py above, which targets specific
+  # shared-object scenarios by hand. -n 0 so the threads share one interpreter.
+  #
+  # --parallel-threads=2 matches CuPy's free-threading CI (the closest GPU
+  # precedent); a small fixed count bounds GPU-memory pressure from concurrent
+  # kernels and stays reproducible across runners, unlike =auto (the runner's
+  # logical-core count).
+  #
+  # Fail fast if the interpreter is not actually GIL-free (wrong build /
+  # PYTHON_GIL=1): pytest-run-parallel does NOT catch a GIL that is enabled from
+  # the start -- it would run threads GIL-serialized and pass vacuously. (A GIL
+  # *re-enabled mid-run* by a non-free-threaded import IS caught by the plugin,
+  # which is why we do not pass --ignore-gil-enabled.)
+  python -c "import sys; assert not sys._is_gil_enabled(), 'GIL is enabled; parallel sweep has no signal'"
+  python -m pytest -n 0 -v --parallel-threads=2 compute/test_no_numba.py
 fi

python/cuda_cccl/tests/compute/conftest.py

@@ -88,14 +88,21 @@ def cuda_stream() -> Generator[Stream, None, None]:
 
 
 @pytest.fixture(scope="function", autouse=True)
-def verify_sass(request, monkeypatch):
+def verify_sass(request):
     if request.node.get_closest_marker("no_verify_sass"):
         return
 
     if not check_ldl_stl_in_sass:
-        print("not checking sass")
         return
 
+    # Pull monkeypatch dynamically rather than as a fixture parameter so this
+    # autouse fixture does not add monkeypatch to every test's static fixture
+    # closure. pytest-run-parallel treats monkeypatch as thread-unsafe based on
+    # that closure, so a parameter here would serialize the entire free-threaded
+    # parallel sweep -- even though this fixture only patches on the opt-in
+    # SASS-check path (check_ldl_stl_in_sass, off by default and in CI).
+    monkeypatch = request.getfixturevalue("monkeypatch")
+
     import cuda.compute._cccl_interop
 
     monkeypatch.setattr(
@@ -124,8 +131,17 @@ def pytest_collection_modifyitems(config, items):
     serialization_skip = pytest.mark.skip(
         reason="serialization not supported on v2 (HostJIT) backend"
     )
+    # Under the pytest-run-parallel sweep (--parallel-threads > 1) skip the
+    # blanket raise_on_numba_import injection: it monkeypatches
+    # builtins.__import__, which pytest-run-parallel treats as thread-unsafe and
+    # would serialize every no_numba test, neutering the sweep.
+    #
+    # getoption returns the int 1 from the argparse default but a *string* for
+    # CLI-passed values ("2", "auto", "1"), so normalize to str before comparing
+    # -- a bare `> 1` would raise TypeError on the "2" string.
+    running_parallel = str(config.getoption("parallel_threads", 1)) != "1"
     for item in items:
-        if item.get_closest_marker("no_numba"):
+        if item.get_closest_marker("no_numba") and not running_parallel:
             if "raise_on_numba_import" not in item.fixturenames:
                 item.fixturenames.append("raise_on_numba_import")
         if USING_V2 and item.get_closest_marker("serialization"):

python/cuda_cccl/tests/compute/test_no_numba.py

@@ -103,7 +103,11 @@ def _raw_negate_i16_op() -> RawOp:
     return _raw_op(source, "no_numba_negate_i16")
 
 
-def test_import_numba_raises():
+def test_import_numba_raises(raise_on_numba_import):
+    # Request the guard explicitly rather than relying on the no_numba
+    # auto-injection (skipped under the parallel sweep, see conftest
+    # pytest_collection_modifyitems): this test exercises the guard directly, and
+    # the monkeypatch it depends on keeps it single-threaded under the sweep.
     with pytest.raises(
         ImportError, match="This test is marked 'no_numba' but attempted to import it"
     ):
@@ -205,9 +209,8 @@ def test_binary_search_explicit_opkind_less(search, side):
     np.testing.assert_array_equal(d_out.copy_to_host(), expected)
 
 
-def test_segmented_reduce_well_known_plus(monkeypatch):
-    monkeypatch.setattr(cuda.compute._cccl_interop, "_check_sass", False)
-
+@pytest.mark.no_verify_sass
+def test_segmented_reduce_well_known_plus():
     h_input = np.asarray([1, 2, 3, 4, 5, 6, 7, 8], dtype=np.uint32)
     h_starts = np.asarray([0, 3, 5], dtype=np.int32)
     h_ends = np.asarray([3, 5, 8], dtype=np.int32)
@@ -288,11 +291,8 @@ def test_segmented_sort_keys():
     np.testing.assert_array_equal(d_output.copy_to_host(), expected)
 
 
-def test_unique_by_key_well_known_equal_to(monkeypatch):
-    cc_major, _ = cuda.compute._cccl_interop.CudaDevice().compute_capability
-    if cc_major >= 9:
-        monkeypatch.setattr(cuda.compute._cccl_interop, "_check_sass", False)
-
+@pytest.mark.no_verify_sass
+def test_unique_by_key_well_known_equal_to():
     h_keys = np.asarray([1, 1, 2, 2, 2, 3, 4, 4], dtype=np.int16)
     h_values = np.asarray([10, 11, 20, 21, 22, 30, 40, 41], dtype=np.int8)
     d_keys = DeviceArray.from_numpy(h_keys)