diff --git a/paddle/gserver/activations/ActivationFunction.cpp b/paddle/gserver/activations/ActivationFunction.cpp
index 81cc3c890b6d4ad048e4edc03208c85778244078..5de2170877ed6f6c70c5617918ad2c4e3b3ed2ee 100644
--- a/paddle/gserver/activations/ActivationFunction.cpp
+++ b/paddle/gserver/activations/ActivationFunction.cpp
@@ -186,7 +186,10 @@ Error __must_check forward(Argument& act) {
                                     useGpu(act.deviceId));
   }
 
-  auto starts = act.sequenceStartPositions->getVector(useGpu(act.deviceId));
+  auto starts =
+      act.hasSubseq()
+          ? act.subSequenceStartPositions->getVector(useGpu(act.deviceId))
+          : act.sequenceStartPositions->getVector(useGpu(act.deviceId));
   act.value->sequenceSoftmax(*act.value, *starts);
   return Error();
 }
@@ -197,8 +200,9 @@ Error __must_check backward(Argument& act) {
         "Input width for each timestep of sequence softmax should be 1");
   }
 
-  size_t numSequences = act.getNumSequences();
-  const int* starts = act.sequenceStartPositions->getData(false);
+  size_t numSequences =
+      act.hasSubseq() ? act.getNumSubSequences() : act.getNumSequences();
+  const int* starts = act.getCpuStartPositions();
 
   for (size_t i = 0; i < numSequences; ++i) {
     // TODO(Dangqingqing) optimization for GPU
diff --git a/paddle/gserver/tests/test_ActivationGrad.cpp b/paddle/gserver/tests/test_ActivationGrad.cpp
index b201ba8a5a4146ab28cd96454f434f889d72a968..de93972a5880518dfbfb9f8582e17c594e54b9b8 100644
--- a/paddle/gserver/tests/test_ActivationGrad.cpp
+++ b/paddle/gserver/tests/test_ActivationGrad.cpp
@@ -57,6 +57,39 @@ TEST(Activation, activation) {
   }
 }
 
+void testSequenceSoftmaxAct(bool hasSubseq) {
+  LOG(INFO) << "test activation: sequence softmax";
+
+  const size_t size = 1;
+  TestConfig config;
+  config.biasSize = 0;
+  config.layerConfig.set_type("addto");
+  config.layerConfig.set_size(size);
+  config.layerConfig.set_active_type("sequence_softmax");
+  config.inputDefs.push_back(
+      {hasSubseq ? INPUT_HASSUB_SEQUENCE_DATA : INPUT_SEQUENCE_DATA,
+       "layer_0",
+       1,
+       0});
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config,
+                  "sequence_softmax",
+                  100,
+                  /* trans= */ false,
+                  useGpu,
+                  /* useWeight */ true);
+  }
+}
+
+TEST(SequenceSoftmaxActivation, activation) {
+  for (auto hasSubseq : {false, true}) {
+    LOG(INFO) << "hasSubseq = " << hasSubseq;
+    testSequenceSoftmaxAct(hasSubseq);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);