diff --git a/paddle/fluid/operators/split_op.cc b/paddle/fluid/operators/split_op.cc
index a05582ae09e16ee17194d299d713d321f28ccace..a43bad878179d02c41d8c8bcd6b43eaffaa6e9a2 100644
--- a/paddle/fluid/operators/split_op.cc
+++ b/paddle/fluid/operators/split_op.cc
@@ -39,14 +39,22 @@ class SplitOp : public framework::OperatorWithKernel {
 
     if (num > 0) {
       int64_t in_axis_dim = in_dims[axis];
-      PADDLE_ENFORCE_EQ(in_axis_dim % num, 0,
-                        "tensor split does not result"
-                        " in an equal division");
-      size_t out_axis_dim = in_axis_dim / num;
-      for (size_t i = 0; i < outs_number; ++i) {
-        auto dim = in_dims;
-        dim[axis] = out_axis_dim;
-        outs_dims.push_back(dim);
+      if (ctx->IsRuntime() || in_axis_dim > 0) {
+        PADDLE_ENFORCE_EQ(in_axis_dim % num, 0,
+                          "tensor split does not result"
+                          " in an equal division");
+        size_t out_axis_dim = in_axis_dim / num;
+        for (size_t i = 0; i < outs_number; ++i) {
+          auto dim = in_dims;
+          dim[axis] = out_axis_dim;
+          outs_dims.push_back(dim);
+        }
+      } else {
+        for (size_t i = 0; i < outs_number; ++i) {
+          auto dim = in_dims;
+          dim[axis] = -1;
+          outs_dims.push_back(dim);
+        }
       }
     } else if (sections.size() > 0) {
       PADDLE_ENFORCE_EQ(sections.size(), outs_number,