diff --git a/test/registered/unit/mem_cache/test_cp_shared_kv_runtime.py b/test/registered/unit/mem_cache/test_cp_shared_kv_runtime.py
index d6a7b0850..48d69e6b1 100644
--- a/test/registered/unit/mem_cache/test_cp_shared_kv_runtime.py
+++ b/test/registered/unit/mem_cache/test_cp_shared_kv_runtime.py
@@ -6,6 +6,20 @@ from unittest.mock import patch
 
 import torch
 
+# Prefer the real compiled sgl_kernel when it is available (e.g. on a GPU box).
+# The CPU-CI stubs below register torch.library ops and a sys.modules module stub
+# at import time. If this module is imported before the real sgl_kernel loads, the
+# empty stub permanently shadows it process-wide, so a later test that needs the
+# genuine kernel (e.g. fast_topk_transform_ragged_fused) calls a None-returning
+# lambda. Loading the real module first makes the FRAGMENT defines hit the
+# already-registered path and keeps sys.modules["sgl_kernel"] pointing at the real
+# module (setattr only fills genuinely-missing names). On CPU-CI the import fails
+# and the stubs are built exactly as before.
+try:
+    import sgl_kernel  # noqa: F401
+except Exception:
+    pass
+
 _sgl_kernel_lib = torch.library.Library("sgl_kernel", "FRAGMENT")