diff --git a/ci/_utils.sh b/ci/_utils.sh
index d09776b93a..8f634d4655 100644
--- a/ci/_utils.sh
+++ b/ci/_utils.sh
@@ -303,11 +303,13 @@ get_ctest_junitxml() {
 
 check_test_filter() {
     test -z "$TEST_FILTER" && return 0
+    set -f # disable globbing so patterns in masks are used as-is 
     for _tf in $TEST_FILTER; do
         case "$1" in
-        $_tf) return 0
+        $_tf) set +f; return 0
         esac
     done
+    set +f
     return 1
 }
 
diff --git a/tests/pytorch/triton_kernels/conftest.py b/tests/pytorch/triton_kernels/conftest.py
index 9cbf0bffdd..78e354478c 100644
--- a/tests/pytorch/triton_kernels/conftest.py
+++ b/tests/pytorch/triton_kernels/conftest.py
@@ -35,14 +35,27 @@
 import torch
 
 
+_force_exit = False
+_exitstatus = 0
+
 @pytest.hookimpl(trylast=True)
 def pytest_sessionfinish(session, exitstatus):
+    global _force_exit, _exitstatus
+
     # Only ROCm hits the hsa_shut_down teardown segfault; leave CUDA/CPU exit
     # semantics (and their normal atexit cleanup) untouched.
-    if getattr(torch.version, "hip", None) is None:
-        return
-    # trylast ensures the junitxml plugin and te_ci_result_sink have already
-    # written their reports in this same hook before we hard-exit.
-    sys.stdout.flush()
-    sys.stderr.flush()
-    os._exit(0 if exitstatus == 0 else int(exitstatus))
+    if getattr(torch.version, "hip", None) is not None:
+        _force_exit = True
+        _exitstatus = int(exitstatus)
+
+# trylast ensures the junitxml plugin and te_ci_result_sink have already
+# written their reports in this same hook before we hard-exit.
+# unconfigure is the last hook called before pytest exits
+@pytest.hookimpl(trylast=True)
+def pytest_unconfigure(config):
+    global _force_exit, _exitstatus
+
+    if _force_exit:
+        sys.stdout.flush()
+        sys.stderr.flush()
+        os._exit(_exitstatus)