diff --git a/paddle/fluid/framework/details/memory_optimize_pass.cc b/paddle/fluid/framework/details/memory_optimize_pass.cc
index c426059a6a6ea640c3c13a667c42d56fbe4fb7c3..fabcd2ecd2b54961c80c40fbcf5dbeb9fe294684 100644
--- a/paddle/fluid/framework/details/memory_optimize_pass.cc
+++ b/paddle/fluid/framework/details/memory_optimize_pass.cc
@@ -128,7 +128,7 @@ std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
       }
     }
   }
-  graph->ResolveHazard(var_nodes_);
+  // graph->ResolveHazard(var_nodes_);
 
   return graph;
 }
@@ -324,6 +324,32 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
   }
 }
 
+void MemoryOptimizePass::ClearControlDepVars(ir::Graph* graph) const {
+  for (auto& op : graph->Nodes()) {
+    if (!op->IsOp()) continue;
+    {
+      auto& nodes = op->inputs;
+      nodes.erase(
+          std::remove_if(nodes.begin(), nodes.end(),
+                         [&](ir::Node* var) { return var->IsCtrlVar(); }),
+          nodes.end());
+    }
+    {
+      auto& nodes = op->outputs;
+      nodes.erase(
+          std::remove_if(nodes.begin(), nodes.end(),
+                         [&](ir::Node* var) { return var->IsCtrlVar(); }),
+          nodes.end());
+    }
+  }
+
+  for (auto& node : graph->Nodes()) {
+    if (node->IsCtrlVar()) {
+      graph->RemoveNode(node);
+    }
+  }
+}
+
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/fluid/framework/details/memory_optimize_pass.h b/paddle/fluid/framework/details/memory_optimize_pass.h
index 593ffc10fc99d26b1ee9174ceef081581126e7e8..f5d188101ffe60fb190702f15753491535f68a3a 100644
--- a/paddle/fluid/framework/details/memory_optimize_pass.h
+++ b/paddle/fluid/framework/details/memory_optimize_pass.h
@@ -48,6 +48,7 @@ class MemoryOptimizePass : public ir::Pass {
   void RenameVarInGraphNode(const std::string& var,
                             const std::string& cache_var, size_t idx,
                             ir::Graph* graph) const;
+  void ClearControlDepVars(ir::Graph* graph) const;
 
   void SubGraphOptimize(OpDesc* op_desc) const;
   // 1. scan op with subblock and collect the output/input vars.
diff --git a/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py b/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py
index 03471a4432f2b6bf6220e79e99aa506628b1535b..c1fb53ecf52d953fa470998c120930b2bec6325b 100644
--- a/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py
+++ b/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py
@@ -121,6 +121,8 @@ class TestMNIST(TestParallelExecutorBase):
                 regularization=fluid.regularizer.L2Decay(1e-6))
             return optimizer
 
+        # NOTE(dzh):
+        # need to make it compatible with elewise fuse act
         not_fuse_op_first_loss, not_fuse_op_last_loss = self.check_network_convergence(
             model,
             feed_dict={"image": img,
@@ -128,6 +130,7 @@ class TestMNIST(TestParallelExecutorBase):
             use_cuda=use_cuda,
             fuse_elewise_add_act_ops=False,
             memory_opt=False,
+            use_ir_memory_optimize=False,
             optimizer=_optimizer)
         fuse_op_first_loss, fuse_op_last_loss = self.check_network_convergence(
             model,
@@ -136,6 +139,7 @@ class TestMNIST(TestParallelExecutorBase):
             use_cuda=use_cuda,
             fuse_elewise_add_act_ops=True,
             memory_opt=False,
+            use_ir_memory_optimize=False,
             optimizer=_optimizer)
 
         for loss in zip(not_fuse_op_first_loss, fuse_op_first_loss):