diff --git a/imperative/src/impl/ops/utility.cpp b/imperative/src/impl/ops/utility.cpp
index 8c68ad741fd306de22f62143a4f60966ef47c92c..21104a726890b2b25c97a777a0152281e8a7e93b 100644
--- a/imperative/src/impl/ops/utility.cpp
+++ b/imperative/src/impl/ops/utility.cpp
@@ -288,6 +288,7 @@ ComputingGraphHolder& get_computing_graph(std::shared_ptr<OpDef> compiled_op, Sm
         cg_holder.graph->options().async_exec_level = 0;
         cg_holder.graph->options().graph_opt_level = compiled_op->cast_final_safe<CompiledOp>().gopt_level;
         cg_holder.graph->options().enable_var_mem_defragment = false;
+        cg_holder.graph->options().comp_seq_sync_device = false;
         cg_holder.graph->set_device_memory_allocator(cg_holder.allocator);
         // cg_holder.graph->options().graph_opt.jit = 2;
         VarNodeArray input_vars;
diff --git a/src/core/impl/graph/cg_impl_seq.cpp b/src/core/impl/graph/cg_impl_seq.cpp
index 1d5444d654cc521bbb6d55e9d57c23ce10c1d6e6..b5fa89e363997e00682f779b0c25841c4b2456cb 100644
--- a/src/core/impl/graph/cg_impl_seq.cpp
+++ b/src/core/impl/graph/cg_impl_seq.cpp
@@ -385,21 +385,27 @@ void ComputingGraphImpl::ComputingSequence::do_wait(bool explicit_user_wait) {
         }
     }
 
-    for (auto cn : m_used_comp_node) {
-        m_event_end.at(cn)->host_wait();
+    bool sync_device = m_owner_graph->options().comp_seq_sync_device;
+
+    if (sync_device) {
+        for (auto cn : m_used_comp_node) {
+            m_event_end.at(cn)->host_wait();
+        }
     }
     m_wait_finished = true;
 #if MGB_NEED_MEGDNN_ASYNC_ERROR
     // FIXME: It CAN NOT work well if more than one ComputingSequnces has been
     // executed on the same compnode and got AsyncError concurrently, because
     // only the first async error on each comp_node would be recorded.
-    for (auto&& cn : m_used_comp_node) {
-        auto error = cn.check_async_error();
-        if (error) {
-            static_cast<const OperatorNodeExcExtraInfo*>(error->extra_info())
-                    ->opr()
-                    ->owner_graph()
-                    ->record_async_error(std::move(error));
+    if (sync_device) {
+        for (auto&& cn : m_used_comp_node) {
+            auto error = cn.check_async_error();
+            if (error) {
+                static_cast<const OperatorNodeExcExtraInfo*>(error->extra_info())
+                        ->opr()
+                        ->owner_graph()
+                        ->record_async_error(std::move(error));
+            }
         }
     }
 #endif
diff --git a/src/core/include/megbrain/graph/cg.h b/src/core/include/megbrain/graph/cg.h
index bbca52052d440113beb8523c3d59f9ee73b8949a..d2de7c25d9f885beaef75212f52e8e060170d6ae 100644
--- a/src/core/include/megbrain/graph/cg.h
+++ b/src/core/include/megbrain/graph/cg.h
@@ -520,6 +520,9 @@ class ComputingGraph : public std::enable_shared_from_this<ComputingGraph>,
              */
             bool no_force_inplace = false;
 
+            //! whether to sync comp_node when waiting computing sequence
+            bool comp_seq_sync_device = true;
+
             //! add extra deps for the comp seq if a specific var is dependent
             ThinHashMap<VarNode*, VarNodeArray> extra_vardeps;