diff --git a/sgl-kernel/src/sgl-kernel/__init__.py b/sgl-kernel/src/sgl-kernel/__init__.py
index a3d35072d..ff41db8e4 100644
--- a/sgl-kernel/src/sgl-kernel/__init__.py
+++ b/sgl-kernel/src/sgl-kernel/__init__.py
@@ -1,3 +1,12 @@
+import ctypes
+import os
+
+if os.path.exists("/usr/local/cuda/targets/x86_64-linux/lib/libcudart.so.12"):
+    ctypes.CDLL(
+        "/usr/local/cuda/targets/x86_64-linux/lib/libcudart.so.12",
+        mode=ctypes.RTLD_GLOBAL,
+    )
+
 from sgl_kernel.ops import (
     apply_rope_with_cos_sin_cache_inplace,
     bmm_fp8,