diff --git a/paddle/fluid/inference/tensorrt/convert/split_op.cc b/paddle/fluid/inference/tensorrt/convert/split_op.cc
index 871354267e936bf0a31c44efe59e2102872c1ad4..ae5b1b98060a4e73b2d1761d4edafb152f364070 100644
--- a/paddle/fluid/inference/tensorrt/convert/split_op.cc
+++ b/paddle/fluid/inference/tensorrt/convert/split_op.cc
@@ -19,9 +19,6 @@ namespace paddle {
 namespace inference {
 namespace tensorrt {
 
-/*
- * SplitOp.
- */
 class SplitOpConverter : public OpConverter {
  public:
   void operator()(const framework::proto::OpDesc& op,
@@ -40,15 +37,11 @@ class SplitOpConverter : public OpConverter {
     int axis = boost::get<int>(op_desc.GetAttr("axis"));
     std::vector<int> output_lengths =
         boost::get<std::vector<int>>(op_desc.GetAttr("sections"));
-    // PADDLE_ENFORCE(axis != 0);
-    if (axis < 0) {
-      axis += input_dims.nbDims;
-    } else {
-      axis -= 1;
-    }
+    // split on batch is not supported in TensorRT
+    PADDLE_ENFORCE(axis != 0);
+    axis += (axis < 0) ? input_dims.nbDims : -1;
 
     PADDLE_ENFORCE(output_lengths.size() == output_num);
-    //
     plugin::SplitPlugin* plugin = new plugin::SplitPlugin(axis, output_lengths);
     nvinfer1::IPluginLayer* layer =
         engine_->AddPlugin(&input, input_num, plugin);
diff --git a/paddle/fluid/inference/tensorrt/convert/test_split_op.cc b/paddle/fluid/inference/tensorrt/convert/test_split_op.cc
index 23909378dde0678106313bc94207a40413fb353d..5aacc5c600dd1371e3865adc888bb8e24640e7d9 100644
--- a/paddle/fluid/inference/tensorrt/convert/test_split_op.cc
+++ b/paddle/fluid/inference/tensorrt/convert/test_split_op.cc
@@ -59,21 +59,54 @@ void TensorRTSplitTest(const std::vector<int> &in_shape,
   validator.Execute(BatchSize);
 }
 
-TEST(split_op, test_same_shape_batch1) {
+// batch = 0, axis = 1, same shape
+TEST(split_op, test_same_shape_axis1_batch1) {
   TensorRTSplitTest<1, 1>({4, 2, 2}, {2, 2});
 }
-
-TEST(split_op, test_different_shape_batch1) {
+// batch = 0, axis = 1, different shape
+TEST(split_op, test_different_shape_axis1_batch1) {
   TensorRTSplitTest<1, 1>({3, 2, 2}, {2, 1});
 }
-
-TEST(split_op, test_same_shape_batch10) {
+// batch = 10, axis = 1, same shape
+TEST(split_op, test_same_shape_axis1_batch10) {
   TensorRTSplitTest<10, 1>({4, 2, 2}, {2, 2});
 }
-
-TEST(split_op, test_different_shape_batch10) {
+// batch = 10, axis = 1, different shape
+TEST(split_op, test_different_shape_axis1_batch10) {
   TensorRTSplitTest<10, 1>({3, 2, 2}, {2, 1});
 }
+// batch = 0, axis = 2, same shape
+TEST(split_op, test_same_shape_axis2_batch1) {
+  TensorRTSplitTest<1, 2>({3, 4, 2}, {2, 2});
+}
+// batch = 0, axis = 2, different shape
+TEST(split_op, test_different_shape_axis2_batch1) {
+  TensorRTSplitTest<1, 2>({3, 3, 2}, {2, 1});
+}
+// batch = 10, axis = 2, same shape
+TEST(split_op, test_same_shape_axis2_batch10) {
+  TensorRTSplitTest<10, 2>({3, 4, 2}, {2, 2});
+}
+// batch = 10, axis = 2, different shape
+TEST(split_op, test_different_shape_axis2_batch10) {
+  TensorRTSplitTest<10, 2>({3, 3, 2}, {2, 1});
+}
+// batch = 0, axis = 3, same shape
+TEST(split_op, test_same_shape_axis3_batch1) {
+  TensorRTSplitTest<1, 3>({3, 2, 4}, {2, 2});
+}
+// batch = 0, axis = 3, different shape
+TEST(split_op, test_different_shape_axis3_batch1) {
+  TensorRTSplitTest<1, 3>({3, 2, 3}, {2, 1});
+}
+// batch = 10, axis = 3, same shape
+TEST(split_op, test_same_shape_axis3_batch10) {
+  TensorRTSplitTest<10, 3>({3, 2, 4}, {2, 2});
+}
+// batch = 10, axis = 3, different shape
+TEST(split_op, test_different_shape_axis3_batch10) {
+  TensorRTSplitTest<10, 3>({3, 2, 3}, {2, 1});
+}
 
 }  // namespace tensorrt
 }  // namespace inference