diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
index 9f48b8cb9e72e7b826156502d2b89bd576388108..e4bbcabea71ec9aa7ebff9d5371a0f49e3f6f7b7 100644
--- a/paddle/fluid/framework/operator.cc
+++ b/paddle/fluid/framework/operator.cc
@@ -926,12 +926,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
     dev_ctx = pool.Get(expected_kernel_key.place_);
   }
 
-  // If Op has attribute all_kernels_must_compute_runtime_shape,
-  // all the kernels of this Op would compute runtime shape,
-  // and skip infershape in runtime for speedup.
-  // TODO(luotao): Note that it is a temporal attribute, after all ops
-  // implement computing runtime shape, this attribute would be deleted.
-  if (!HasAttr("all_kernels_must_compute_runtime_shape")) {
+  if (!HasAttr(kAllKernelsMustComputeRuntimeShape)) {
     RuntimeInferShapeContext infer_shape_ctx(*this, exec_scope, ctx);
     this->InferShape(&infer_shape_ctx);
   }
diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h
index 55629636a816982c4debe4b5b7138558ac309eb5..822bf5c9ceaa31e1283fa3cf1dbe42a43894a5dd 100644
--- a/paddle/fluid/framework/operator.h
+++ b/paddle/fluid/framework/operator.h
@@ -62,6 +62,15 @@ constexpr char kZeroVarSuffix[] = "@ZERO";
 /// Variables with this suffix are the new Gradient.
 constexpr char kNewGradSuffix[] = "@NEWGRAD@";
 
+/// If an Op has this attribute, all its kernels should calculate output
+/// variable's shape in the corresponding Compute() function. And
+/// OperatorWithKernel::RunImpl() would skip call this Op's InferShape()
+/// function in its runtime for speedup.
+/// TODO(luotao): Note that this temporal attribute would be deleted after all
+/// ops contain it.
+constexpr char kAllKernelsMustComputeRuntimeShape[] =
+    "@ALL_KERNELS_MUST_COMPUTE_RUNTIME_SHAPE@";
+
 // define some kernel priority
 /* Define multiple kernel type fallback order*/
 extern std::vector<std::tuple<platform::Place, LibraryType>> kKernelPriority;
diff --git a/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.cc b/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.cc
index 17a81d3e8805a5ed9fc9bb5310227df78eefa056..a0026427e2514735711f7eba26fcf861cb498d5e 100644
--- a/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.cc
+++ b/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.cc
@@ -88,13 +88,7 @@ class FusedEmbeddingSeqPoolOpMaker : public framework::OpProtoAndCheckerMaker {
                   "(boolean, default false) "
                   "Sparse update.")
         .SetDefault(false);
-    AddAttr<bool>(
-        "all_kernels_must_compute_runtime_shape",
-        "(boolean, default true) "
-        "An attribute to speed up OperatorWithKernel::RunImpl."
-        "If true, all the kernels of this Op would compute runtime "
-        "shape, but skip infershape in runtime. Note that it is a temporal "
-        "attribute, please do DOT set it in python layer.")
+    AddAttr<bool>(framework::kAllKernelsMustComputeRuntimeShape, "")
         .SetDefault(true);
     AddComment(R"DOC(
 FusedEmbeddingSeqPool Operator.
diff --git a/paddle/fluid/operators/hash_op.cc b/paddle/fluid/operators/hash_op.cc
index b39eba081ec32ed50ac78a0ea5b3aff3ed4045d2..f6395fb32feac175976cb96e1c0bee7347cb3ea8 100644
--- a/paddle/fluid/operators/hash_op.cc
+++ b/paddle/fluid/operators/hash_op.cc
@@ -54,13 +54,7 @@ $$Out = scale * X$$
 )DOC");
     AddAttr<int>("num_hash", "").SetDefault(1);
     AddAttr<int>("mod_by", "").SetDefault(100000);
-    AddAttr<bool>(
-        "all_kernels_must_compute_runtime_shape",
-        "(boolean, default true) "
-        "An attribute to speed up OperatorWithKernel::RunImpl."
-        "If true, all the kernels of this Op would compute runtime "
-        "shape, but skip infershape in runtime. Note that it is a temporal "
-        "attribute, please do DOT set it in python layer.")
+    AddAttr<bool>(framework::kAllKernelsMustComputeRuntimeShape, "")
         .SetDefault(true);
   }
 };
diff --git a/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc b/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc
index 63e95e8654412048ff54fc6407984949152f58d4..f357c9c08d042b69259f229955922f2f11b52c63 100644
--- a/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc
+++ b/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc
@@ -59,13 +59,7 @@ class SequenceEnumerateOpMaker : public framework::OpProtoAndCheckerMaker {
         });
     AddAttr<int>("pad_value", "(int) The enumerate sequence padding value.")
         .SetDefault(0);
-    AddAttr<bool>(
-        "all_kernels_must_compute_runtime_shape",
-        "(boolean, default true) "
-        "An attribute to speed up OperatorWithKernel::RunImpl."
-        "If true, all the kernels of this Op would compute runtime "
-        "shape, but skip infershape in runtime. Note that it is a temporal "
-        "attribute, please do DOT set it in python layer.")
+    AddAttr<bool>(framework::kAllKernelsMustComputeRuntimeShape, "")
         .SetDefault(true);
     AddComment(R"DOC(
 Sequence Enumerate Operator.