diff --git a/imperative/python/src/ops.cpp b/imperative/python/src/ops.cpp
index e151fadb..22ed862e 100644
--- a/imperative/python/src/ops.cpp
+++ b/imperative/python/src/ops.cpp
@@ -467,7 +467,9 @@ void init_ops(py::module m) {
     m.def("delete_rng_handle", [](size_t handle){
         // RNG op might execute after handle released due to async dispatch, so
         // we need sync before delete a handle to avoid memory leak or use-after-free
-        python::interpreter_for_py->sync();
+        if(python::interpreter_for_py->check_available()){
+            python::interpreter_for_py->sync();
+        }
         mgb::CompNode::sync_all();
         py_task_q.wait_all_task_finish();
         rng::delete_handle(handle);
diff --git a/imperative/src/impl/interpreter/interpreter_impl.h b/imperative/src/impl/interpreter/interpreter_impl.h
index 00fcf8e8..14b3b2de 100644
--- a/imperative/src/impl/interpreter/interpreter_impl.h
+++ b/imperative/src/impl/interpreter/interpreter_impl.h
@@ -60,6 +60,7 @@ struct ChannelImpl : Interpreter::Channel {
 
     DeviceTensorND get_dev_tensor(Handle) override;
 
+    bool check_available() override;
     void sync() override;
     void close() override;
 
@@ -112,8 +113,6 @@ private:
         const SmallVector<LogicalTensorDesc>& input_descs,
         SmallVector<Handle>* outputs);
 
-    bool check_available();
-
     void push_scope(std::string, State&);
     void pop_scope(std::string, State&);
 
diff --git a/imperative/src/include/megbrain/imperative/interpreter.h b/imperative/src/include/megbrain/imperative/interpreter.h
index a4f29ff4..f7f5028f 100644
--- a/imperative/src/include/megbrain/imperative/interpreter.h
+++ b/imperative/src/include/megbrain/imperative/interpreter.h
@@ -41,6 +41,7 @@ struct Interpreter {
 
         virtual DeviceTensorND get_dev_tensor(Handle) = 0;
 
+        virtual bool check_available() = 0;
         virtual void sync() = 0;
         virtual void close() = 0;