diff --git a/src/core/impl/graph/var_node.cpp b/src/core/impl/graph/var_node.cpp
index 6cfae871ae6b5c9bf2c3b71b3c6e9f0d440c7808..e6a5a1ff594a9b6b79f9460d2c6f2b502e015ddc 100644
--- a/src/core/impl/graph/var_node.cpp
+++ b/src/core/impl/graph/var_node.cpp
@@ -307,7 +307,7 @@ VarNode& VarNode::shape(const TensorShape &shape) {
     return *this;
 }
 
-VarNode& VarNode::shape_alloc(const TensorShape &shape) {
+VarNode& VarNode::shape_alloc(const TensorShape &shape, size_t size_req) {
     mgb_assert(shape.ndim, "got empty shape in shape_alloc: "
             "var=%s owner_opr=%s{%s}", cname(), owner_opr()->cname(),
             owner_opr()->dyn_typeinfo()->name);
@@ -316,7 +316,7 @@ VarNode& VarNode::shape_alloc(const TensorShape &shape) {
                 " NO_SYS_MEM_ALLOC flag; actual var: %s",
                 cg::dump_var_info({this}).c_str());
     ComputingGraphImpl::downcast(owner_graph())
-        ->var_node_mem_manager().var_alloc_with_shape(this, shape);
+        ->var_node_mem_manager().var_alloc_with_shape(this, shape, size_req);
     return *this;
 }
 
diff --git a/src/core/impl/graph/var_node_mem_mgr.cpp b/src/core/impl/graph/var_node_mem_mgr.cpp
index 381436b08c1f7cb09aee0bb2673e1d8df7c8f211..3e8bc615e901992162676f695953e8650905bbd3 100644
--- a/src/core/impl/graph/var_node_mem_mgr.cpp
+++ b/src/core/impl/graph/var_node_mem_mgr.cpp
@@ -1239,13 +1239,18 @@ void VarNodeMemManager::make_dev_tensor_from_mem_plan_single(
 }
 
 void VarNodeMemManager::var_alloc_with_shape(VarNode* var,
-                                             const TensorShape& shape) {
+                                             const TensorShape& shape,
+                                             size_t size_req) {
     mgb_assert(var->format().is_default(),
                "dynamic shape is currently only supported for var with "
                "default format; got %s",
                var->format().to_string().c_str());
     var->shape(shape);
-    auto size_req = var->dtype().size(shape.total_nr_elems());
+    if (size_req != 0) {
+        mgb_assert(var->dtype().size(shape.total_nr_elems()) <= size_req);
+    } else {
+        size_req = var->dtype().size(shape.total_nr_elems());
+    }
 
     auto&& mplan = var->m_mem_plan;
     if (!mplan.valid() || mplan.chunk().owner_var != var)
diff --git a/src/core/impl/graph/var_node_mem_mgr.h b/src/core/impl/graph/var_node_mem_mgr.h
index 2f2e99717f288ee174e9853de2685ba361e1c33f..00e19cd638572902d53428c80dc9162582e12682 100644
--- a/src/core/impl/graph/var_node_mem_mgr.h
+++ b/src/core/impl/graph/var_node_mem_mgr.h
@@ -294,7 +294,13 @@ class VarNodeMemManager {
         void add_layout_constraint_level(
                 VarNode *dest, LayoutConstraintLevel level);
 
-        void var_alloc_with_shape(VarNode *var, const TensorShape &shape);
+        /**
+         * \brief alloc var memory with shape.
+         *
+         * Alloc memory of size_seq if size_req != 0.
+         */
+        void var_alloc_with_shape(VarNode* var, const TensorShape& shape,
+                                  size_t size_req = 0);
 
         /*!
          * \brief initialize mem plan for a single var
diff --git a/src/core/include/megbrain/graph/var_node.h b/src/core/include/megbrain/graph/var_node.h
index f9a5d4d262fb47f2039b587fdf6457e2d5d169dd..31e4fef2bcda8aaafbfa5f8ff1a9e4368b808966 100644
--- a/src/core/include/megbrain/graph/var_node.h
+++ b/src/core/include/megbrain/graph/var_node.h
@@ -462,8 +462,10 @@ class VarNode final: public GraphNodeBase {
          * this var must have NO_SYS_MEM_ALLOC flag; if shape does not increase
          * and original tensor storage is valid, it is guaranteed that old data
          * would be retained.
+         *
+         * \warning Alloc size_req memory if size_req != 0.
          */
-        VarNode& shape_alloc(const TensorShape &shape);
+        VarNode& shape_alloc(const TensorShape &shape, size_t size_req = 0);
 
         /*!
          * \brief directly reset device tensor from another var