diff --git a/paddle/fluid/operators/hierarchical_sigmoid_op.h b/paddle/fluid/operators/hierarchical_sigmoid_op.h
index b47bf49ecb7c0f56ec3cffc02bef8c034417555f..1a7ca963010112bbcab69f1ceeb9cb8d19ca9b9e 100644
--- a/paddle/fluid/operators/hierarchical_sigmoid_op.h
+++ b/paddle/fluid/operators/hierarchical_sigmoid_op.h
@@ -40,8 +40,9 @@ using platform::Transform;
 
 static std::vector<int64_t> PathToRows(const framework::LoDTensor& path) {
   std::set<int64_t> rows;
+  const int64_t* paths = path.data<int64_t>();
   for (int64_t i = 0; i < path.numel(); ++i) {
-    int64_t row = path.data<int64_t>()[i];
+    int64_t row = paths[i];
     if (row < 0) {
       continue;
     }
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
index 861bc3202615c419cb8325e167d9925e6d860fa7..6379031ee49a7ae79fdfa7ac3b045781ee95545e 100644
--- a/python/paddle/fluid/layers/nn.py
+++ b/python/paddle/fluid/layers/nn.py
@@ -5013,9 +5013,10 @@ def nce(input,
     else:
         num_neg_samples = int(num_neg_samples)
 
-    remote_prefetch = False
-    if os.environ.get('PADDLE_ENABLE_REMOTE_PREFETCH'):
-        remote_prefetch = True
+    remote_prefetch = is_sparse
+    print(
+        "With sparse mode, if your models has only small parameter prefetch may cause speed down"
+    )
 
     attrs = {
         'num_total_classes': int(num_total_classes),
@@ -5133,10 +5134,10 @@ def hsigmoid(input,
         pass
 
     weights = None
-    remote_prefetch = False
-    if os.environ.get('PADDLE_ENABLE_REMOTE_PREFETCH'):
-        remote_prefetch = True
-
+    remote_prefetch = is_sparse
+    print(
+        "With sparse mode, if your models has only small parameter prefetch may cause speed down"
+    )
     if not is_custom:
         weights = helper.create_parameter(
             attr=helper.param_attr,