diff --git a/paddle/fluid/framework/details/memory_optimize_helper.cc b/paddle/fluid/framework/details/memory_optimize_helper.cc
index 6126c168ccf4b52ae75648d2dfe48db770639214..db4e805bb692ee44ac50337fae54f8dbfe389e6f 100644
--- a/paddle/fluid/framework/details/memory_optimize_helper.cc
+++ b/paddle/fluid/framework/details/memory_optimize_helper.cc
@@ -129,7 +129,13 @@ size_t NodeSize(const VarDesc& node) {
 }
 
 size_t NodeSize(ir::Node* n) {
-  auto* desc = FindVarDescInBlock(n);
+  VarDesc* desc = nullptr;
+  // some op do not have block pointer
+  if (n->inputs[0]->Op() != nullptr) {
+    desc = FindVarDescInBlock(n);
+  } else {
+    desc = n->Var();
+  }
   return NodeSize(*desc);
 }
 
diff --git a/paddle/fluid/framework/details/memory_optimize_pass.cc b/paddle/fluid/framework/details/memory_optimize_pass.cc
index 93d08649db2673b31faad39e5b6d2aea390a8e08..d45a43d851dc7dff1421a4f8e3316a8b1169e979 100644
--- a/paddle/fluid/framework/details/memory_optimize_pass.cc
+++ b/paddle/fluid/framework/details/memory_optimize_pass.cc
@@ -194,7 +194,8 @@ void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const {
           // effect. Because it is a single op in graph. No need to
           // update the ir nodes.
           sub_op_desc->Rename(var->Name(), cache->Name());
-          if (sub_op_desc->Block()->HasVar(var->Name())) {
+          if (sub_op_desc->Block() != nullptr &&
+              sub_op_desc->Block()->HasVar(var->Name())) {
             sub_op_desc->Block()->RemoveVar(var->Name());
           }
         }