diff --git a/paddle/fluid/operators/pad_constant_like_op.cc b/paddle/fluid/operators/pad_constant_like_op.cc
index 3f827c26fd415c8a3c2295129f413850ea59bef3..920383e6522e6bea0df98fe03c7bc197d9e18f09 100644
--- a/paddle/fluid/operators/pad_constant_like_op.cc
+++ b/paddle/fluid/operators/pad_constant_like_op.cc
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/pad_constant_like_op.h"
+#include <memory>
 
 namespace paddle {
 namespace operators {
@@ -38,8 +39,13 @@ class PadConstantLikeOp : public framework::OperatorWithKernel {
                       "The dimention of X and Y should be the same.");
 
     for (int i = 0; i < x_dim.size(); ++i) {
-      PADDLE_ENFORCE_GE(x_dim[i], y_dim[i]);
+      if ((!ctx->IsRuntime()) && ((x_dim[i] == -1) || (y_dim[i] == -1))) {
+        continue;
+      } else {
+        PADDLE_ENFORCE_GE(x_dim[i], y_dim[i]);
+      }
     }
+
     ctx->SetOutputDim("Out", x_dim);
     ctx->ShareLoD("X", /*->*/ "Out");
   }
@@ -162,7 +168,11 @@ class PadConstantLikeOpGrad : public framework::OperatorWithKernel {
       ctx->ShareLoD("Y", /*->*/ y_grad_name);
 
       for (int i = 0; i < y_dim.size(); ++i) {
-        PADDLE_ENFORCE_GE(dout_dim[i], y_dim[i]);
+        if ((!ctx->IsRuntime()) && ((dout_dim[i] == -1) || (y_dim[i] == -1))) {
+          continue;
+        } else {
+          PADDLE_ENFORCE_GE(dout_dim[i], y_dim[i]);
+        }
       }
     }
   }
diff --git a/paddle/fluid/operators/pad_op.cc b/paddle/fluid/operators/pad_op.cc
index c28106d31273cb54e3974d186296644272d2014c..4b4f43344f6ffff3f7c137edf5fcaf7149fbd72a 100644
--- a/paddle/fluid/operators/pad_op.cc
+++ b/paddle/fluid/operators/pad_op.cc
@@ -36,7 +36,11 @@ class PadOp : public framework::OperatorWithKernel {
                       "of input tensor.");
     std::vector<int64_t> out_dims(x_dim.size());
     for (int i = 0; i < x_dim.size(); ++i) {
-      out_dims[i] = x_dim[i] + paddings[i * 2] + paddings[i * 2 + 1];
+      if ((!ctx->IsRuntime()) && (x_dim[i] == -1)) {
+        out_dims[i] = -1;
+      } else {
+        out_dims[i] = x_dim[i] + paddings[i * 2] + paddings[i * 2 + 1];
+      }
     }
     ctx->SetOutputDim("Out", framework::make_ddim(out_dims));
     if (out_dims[0] == x_dim[0]) {
@@ -100,18 +104,14 @@ class PadOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
   void InferShape(framework::InferShapeContext* ctx) const override {
-    auto dout_dims = ctx->GetInputDim(framework::GradVarName("Out"));
-    auto& paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
-    for (int i = 0; i < dout_dims.size(); ++i) {
-      dout_dims[i] -= (paddings[i * 2] + paddings[i * 2 + 1]);
-    }
-
     auto x_grad_name = framework::GradVarName("X");
     if (ctx->HasOutput(x_grad_name)) {
       auto dout_dims = ctx->GetInputDim(framework::GradVarName("Out"));
       auto& paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
       for (int i = 0; i < dout_dims.size(); ++i) {
-        dout_dims[i] -= (paddings[i * 2] + paddings[i * 2 + 1]);
+        if (ctx->IsRuntime() || (dout_dims[i] != -1)) {
+          dout_dims[i] -= (paddings[i * 2] + paddings[i * 2 + 1]);
+        }
       }
       ctx->SetOutputDim(x_grad_name, dout_dims);
     }