diff --git a/paddle/gserver/layers/SequenceLastInstanceLayer.cpp b/paddle/gserver/layers/SequenceLastInstanceLayer.cpp
index 7ac087663a6c64e4570bd66c3880d1c19b85719f..c70c2b74211814559b8982aae94eed92258444e8 100644
--- a/paddle/gserver/layers/SequenceLastInstanceLayer.cpp
+++ b/paddle/gserver/layers/SequenceLastInstanceLayer.cpp
@@ -72,7 +72,8 @@ bool SequenceLastInstanceLayer::init(const LayerMap& layerMap,
 void SequenceLastInstanceLayer::forward(PassType passType) {
   SequencePoolLayer::forward(passType);
 
-  const int* starts = startPositions_->getData(false);
+  auto starts = (stride_ > 0) ? stridePositions_->getData()
+                              : startPositions_->getData(false);
   MatrixPtr inputValue = getInputValue(0);
   MatrixPtr outputValue = getOutputValue();
 
@@ -82,10 +83,7 @@ void SequenceLastInstanceLayer::forward(PassType passType) {
 
     insId_.clear();
     for (size_t seqId = 0; seqId < newBatchSize_; ++seqId) {
-      int insId = (stride_ > 0)
-                      ? (select_first_ ? stridePositions_[seqId]
-                                       : stridePositions_[seqId + 1] - 1)
-                      : (select_first_ ? starts[seqId] : starts[seqId + 1] - 1);
+      int insId = select_first_ ? starts[seqId] : starts[seqId + 1] - 1;
       insId_.push_back(insId);
 
       outputValue->subMatrix(seqId, 1, tmpDest_)
diff --git a/paddle/gserver/layers/SequencePoolLayer.cpp b/paddle/gserver/layers/SequencePoolLayer.cpp
index 2bf180a043b98fc3a42b4612dd4eef65ae42b5db..f853905103a0e5814d84b04a6dd0eb5ca6beb8d8 100644
--- a/paddle/gserver/layers/SequencePoolLayer.cpp
+++ b/paddle/gserver/layers/SequencePoolLayer.cpp
@@ -70,7 +70,7 @@ void SequencePoolLayer::forward(PassType passType) {
     CHECK_EQ(input.hasSubseq(), 0UL)
         << "sequence stride pooling is not suitable for hasSubseq now";
     output_.poolSequenceWithStride(input, stride_, &stridePositions_);
-    newBatchSize_ = stridePositions_.size() - 1;
+    newBatchSize_ = stridePositions_->getSize() - 1;
   }
 
   resetOutput(newBatchSize_, dim);
diff --git a/paddle/gserver/layers/SequencePoolLayer.h b/paddle/gserver/layers/SequencePoolLayer.h
index 5ca1c1a82b913f2462bb5c227e15555c31590ceb..92d7a841f0c73421e26e5882241f2b0d0e2fba50 100644
--- a/paddle/gserver/layers/SequencePoolLayer.h
+++ b/paddle/gserver/layers/SequencePoolLayer.h
@@ -48,7 +48,7 @@ protected:
   ICpuGpuVectorPtr startPositions_;
   int stride_;
   // store the start position of each stride window
-  std::vector<int> stridePositions_;
+  IVectorPtr stridePositions_;
 
 public:
   explicit SequencePoolLayer(const LayerConfig& config) : Layer(config) {}
diff --git a/paddle/parameter/Argument.cpp b/paddle/parameter/Argument.cpp
index 2657c00ebb3c79c2ec08c4dc59226682fe961fe5..3cc637587bc28ce02218c38489f48df352b1e574 100644
--- a/paddle/parameter/Argument.cpp
+++ b/paddle/parameter/Argument.cpp
@@ -561,7 +561,7 @@ void Argument::degradeSequence(const Argument& input) {
 
 void Argument::poolSequenceWithStride(const Argument& input,
                                       size_t stride,
-                                      std::vector<int>* stridePostions) {
+                                      IVectorPtr* stridePostions) {
   /*
    * If input.sequenceStartPositions = [0, 9, 14, 17, 30] and stride = 5,
    * then sequenceStartPositions = [0, 2, 3, 4, 7],
@@ -577,10 +577,10 @@ void Argument::poolSequenceWithStride(const Argument& input,
   int* tgtBuf = sequenceStartPositions->getMutableData(false);
   // first index of target sequence and stride positions are both 0
   tgtBuf[0] = 0;
-  (*stridePostions).clear();
+  std::vector<int> stridePos;
   for (size_t seqId = 0; seqId < numSequences; ++seqId) {
     size_t seqLength = starts[seqId + 1] - starts[seqId];
-    (*stridePostions).emplace_back(starts[seqId]);
+    stridePos.emplace_back(starts[seqId]);
     if (seqLength == 0) {
       // empty sequence
       tgtBuf[seqId + 1] = tgtBuf[seqId];
@@ -591,12 +591,15 @@ void Argument::poolSequenceWithStride(const Argument& input,
       int size =
           (seqLength % stride) ? seqLength / stride : seqLength / stride - 1;
       for (int i = 0; i < size; i++) {
-        (*stridePostions).emplace_back((*stridePostions).back() + stride);
+        stridePos.emplace_back(stridePos.back() + stride);
       }
     }
   }
-  (*stridePostions).emplace_back(starts[numSequences]);
-  CHECK_EQ((*stridePostions).size() - 1, tgtBuf[numSequences]);
+  stridePos.emplace_back(starts[numSequences]);
+  int size = stridePos.size();
+  CHECK_EQ(size - 1, tgtBuf[numSequences]);
+  IVector::resizeOrCreate(*stridePostions, size, false);
+  (*stridePostions)->copyFrom(stridePos.data(), size);
 }
 
 void Argument::getValueString(
diff --git a/paddle/parameter/Argument.h b/paddle/parameter/Argument.h
index 760029c2fe6ba3417c69c4772e3fafab3632e4b8..95ea90ffc2a604046252add36b0bb2e493b6050f 100644
--- a/paddle/parameter/Argument.h
+++ b/paddle/parameter/Argument.h
@@ -298,7 +298,7 @@ struct Argument {
    */
   void poolSequenceWithStride(const Argument& input,
                               size_t stride,
-                              std::vector<int>* stridePositions);
+                              IVectorPtr* stridePositions);
   /**
    * @brief getValueString will return the argument's output in string. There
    * are several kinds of output. The keys of output dictionary are 'value',
diff --git a/paddle/parameter/tests/test_argument.cpp b/paddle/parameter/tests/test_argument.cpp
index ba17e8a298e80139c23db414c0b77520e99b7272..692bbada10d03f87e591e6e15a5fa0fb0569c7fd 100644
--- a/paddle/parameter/tests/test_argument.cpp
+++ b/paddle/parameter/tests/test_argument.cpp
@@ -27,8 +27,7 @@ TEST(Argument, poolSequenceWithStride) {
   inStart[3] = 17;
   inStart[4] = 30;
 
-  std::vector<int> stridePositions;
-  stridePositions.clear();
+  IVectorPtr stridePositions;
   output.poolSequenceWithStride(input, 5 /* stride */, &stridePositions);
 
   const int* outStart = output.sequenceStartPositions->getData(false);
@@ -38,10 +37,10 @@ TEST(Argument, poolSequenceWithStride) {
   CHECK_EQ(outStart[3], 4);
   CHECK_EQ(outStart[4], 7);
 
-  CHECK_EQ(stridePositions.size(), 8);
+  CHECK_EQ(stridePositions->getSize(), 8);
   int strideResult[] = {0, 5, 9, 14, 17, 22, 27, 30};
   for (int i = 0; i < 8; i++) {
-    CHECK_EQ(stridePositions[i], strideResult[i]);
+    CHECK_EQ(stridePositions->getData()[i], strideResult[i]);
   }
 }
 
diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py
index 542bbbe086ec0e4bfdaa45cc2efef2345c80015f..5f3250e7224fff8dcbf13cace33f10b272064bdc 100755
--- a/python/paddle/trainer_config_helpers/layers.py
+++ b/python/paddle/trainer_config_helpers/layers.py
@@ -1406,7 +1406,6 @@ def first_seq(input,
     and a long sequence will be shorten. Note that for sequence with 
     sub-sequence, stride is default -1 now.
 
-
     The simple usage is:
 
     .. code-block:: python
@@ -1418,6 +1417,8 @@ def first_seq(input,
     :type name: basestring
     :param input: Input layer name.
     :type input: LayerOutput
+    :param stride: parameter of stride window.  
+    :type stride: Int
     :param layer_attr: extra layer attributes.
     :type layer_attr: ExtraLayerAttribute.
     :return: LayerOutput object.