diff --git a/paddle/fluid/operators/math/softmax.cc b/paddle/fluid/operators/math/softmax.cc
index 6300836e500cfee844b3a7193522c06653f5adb0..fa2018178f44ff4e3b14937c1f508fa8a698e20e 100644
--- a/paddle/fluid/operators/math/softmax.cc
+++ b/paddle/fluid/operators/math/softmax.cc
@@ -19,10 +19,10 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template class SoftmaxFunctor<platform::CPUDeviceContext, float,true>;
-template class SoftmaxFunctor<platform::CPUDeviceContext, float,false>;
-template class SoftmaxFunctor<platform::CPUDeviceContext, double,true>;
-template class SoftmaxFunctor<platform::CPUDeviceContext, double,false>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, float, true>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, float, false>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, double, true>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, double, false>;
 template class SoftmaxGradFunctor<platform::CPUDeviceContext, float>;
 template class SoftmaxGradFunctor<platform::CPUDeviceContext, double>;
 
diff --git a/paddle/fluid/operators/math/softmax_impl.h b/paddle/fluid/operators/math/softmax_impl.h
index 6a0a6c2e46d6d0338eb880a9e8a07f8c3cc08493..7cf98f27251db3cfe5e8e295ed21056f6e5a2963 100644
--- a/paddle/fluid/operators/math/softmax_impl.h
+++ b/paddle/fluid/operators/math/softmax_impl.h
@@ -33,9 +33,9 @@ struct ValueClip {
 };
 
 template <typename DeviceContext, typename T, bool is_test>
-void SoftmaxFunctor<DeviceContext, T, is_test>::operator()(const DeviceContext& context,
-                                                  const framework::Tensor* X,
-                                                  framework::Tensor* Y) {
+void SoftmaxFunctor<DeviceContext, T, is_test>::operator()(
+    const DeviceContext& context, const framework::Tensor* X,
+    framework::Tensor* Y) {
   auto logits = EigenMatrix<T>::From(*X);
   auto softmax = EigenMatrix<T>::From(*Y);
 
@@ -67,40 +67,37 @@ void SoftmaxFunctor<DeviceContext, T, is_test>::operator()(const DeviceContext&
 
 template <typename DeviceContext, typename T>
 class SoftmaxFunctor<DeviceContext, T, true> {
-void operator()(const DeviceContext& context,
-                const framework::Tensor* X,
-                framework::Tensor* Y) {
-  auto logits = EigenMatrix<T>::From(*X);
-  auto softmax = EigenMatrix<T>::From(*Y);
-
-  const int kBatchDim = 0;
-  const int kClassDim = 1;
-
-  const int batch_size = logits.dimension(kBatchDim);
-  const int num_classes = logits.dimension(kClassDim);
-
-  Eigen::DSizes<int, 1> along_class(kClassDim);
-  Eigen::DSizes<int, 2> batch_by_one(batch_size, 1);
-  Eigen::DSizes<int, 2> one_by_class(1, num_classes);
-
-  auto shifted_logits = (logits -
-                         logits.maximum(along_class)
-                             .eval()
-                             .reshape(batch_by_one)
-                             .broadcast(one_by_class));
-
-  softmax.device(*context.eigen_device()) = shifted_logits.exp();
-  softmax.device(*context.eigen_device()) = (softmax *
-                                             softmax.sum(along_class)
-                                                 .inverse()
-                                                 .eval()
-                                                 .reshape(batch_by_one)
-                                                 .broadcast(one_by_class));
-}
+  void operator()(const DeviceContext& context, const framework::Tensor* X,
+                  framework::Tensor* Y) {
+    auto logits = EigenMatrix<T>::From(*X);
+    auto softmax = EigenMatrix<T>::From(*Y);
+
+    const int kBatchDim = 0;
+    const int kClassDim = 1;
+
+    const int batch_size = logits.dimension(kBatchDim);
+    const int num_classes = logits.dimension(kClassDim);
+
+    Eigen::DSizes<int, 1> along_class(kClassDim);
+    Eigen::DSizes<int, 2> batch_by_one(batch_size, 1);
+    Eigen::DSizes<int, 2> one_by_class(1, num_classes);
+
+    auto shifted_logits = (logits -
+                           logits.maximum(along_class)
+                               .eval()
+                               .reshape(batch_by_one)
+                               .broadcast(one_by_class));
+
+    softmax.device(*context.eigen_device()) = shifted_logits.exp();
+    softmax.device(*context.eigen_device()) = (softmax *
+                                               softmax.sum(along_class)
+                                                   .inverse()
+                                                   .eval()
+                                                   .reshape(batch_by_one)
+                                                   .broadcast(one_by_class));
+  }
 };
 
-
-
 template <typename DeviceContext, typename T>
 void SoftmaxGradFunctor<DeviceContext, T>::operator()(
     const DeviceContext& context, const framework::Tensor* y,
diff --git a/paddle/fluid/operators/softmax_op.h b/paddle/fluid/operators/softmax_op.h
index 5bc72aac485acceffd90964582d2a8d748f84adc..bcd63eefc7856d529856378a1f8e2537df94fd9b 100644
--- a/paddle/fluid/operators/softmax_op.h
+++ b/paddle/fluid/operators/softmax_op.h
@@ -36,11 +36,11 @@ class SoftmaxKernel : public framework::OpKernel<T> {
     Tensor Out_2d = framework::ReshapeToMatrix(*Out, rank - 1);
 
     const bool is_test = context.Attr<bool>("is_test");
-    if( is_test == true) {
-      math::SoftmaxFunctor<DeviceContext, T,true>()(
+    if (is_test == true) {
+      math::SoftmaxFunctor<DeviceContext, T, true>()(
           context.template device_context<DeviceContext>(), &X_2d, &Out_2d);
     } else {
-      math::SoftmaxFunctor<DeviceContext, T,false>()(
+      math::SoftmaxFunctor<DeviceContext, T, false>()(
           context.template device_context<DeviceContext>(), &X_2d, &Out_2d);
     }
   }
diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.h b/paddle/fluid/operators/softmax_with_cross_entropy_op.h
index 2eec8541c84c00b1c951a9d393d3f54aa97d0971..c0530e3d8bc407ddd6d7bf6e10a715185d0beb1f 100644
--- a/paddle/fluid/operators/softmax_with_cross_entropy_op.h
+++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.h
@@ -42,8 +42,8 @@ class SoftmaxWithCrossEntropyKernel : public framework::OpKernel<T> {
 
     auto& dev_ctx =
         context.template device_context<platform::CPUDeviceContext>();
-    math::SoftmaxFunctor<platform::CPUDeviceContext, T, false>()(dev_ctx, logits,
-                                                          softmax);
+    math::SoftmaxFunctor<platform::CPUDeviceContext, T, false>()(
+        dev_ctx, logits, softmax);
     math::CrossEntropyFunctor<platform::CPUDeviceContext, T>()(
         dev_ctx, loss, softmax, labels, context.Attr<bool>("soft_label"),
         context.Attr<int>("ignore_index"));
diff --git a/python/paddle/fluid/tests/unittests/test_softmax_op.py b/python/paddle/fluid/tests/unittests/test_softmax_op.py
index 40c3135183a128cd9b7324ce27da798fa2d93afd..3bef24430d9aaeb459555312da0c10a5729fd4bc 100644
--- a/python/paddle/fluid/tests/unittests/test_softmax_op.py
+++ b/python/paddle/fluid/tests/unittests/test_softmax_op.py
@@ -35,6 +35,7 @@ class TestSoftmaxOp(OpTest):
         self.op_type = "softmax"
         self.use_cudnn = False
         self.use_mkldnn = False
+        self.is_test = False
         self.dtype = np.float32
         self.init_kernel_type()
         self.shape = self.get_x_shape()
@@ -48,7 +49,8 @@ class TestSoftmaxOp(OpTest):
         self.outputs = {'Out': out}
         self.attrs = {
             'use_cudnn': self.use_cudnn,
-            'use_mkldnn': self.use_mkldnn
+            'use_mkldnn': self.use_mkldnn,
+            'is_test': self.is_test
         }
 
     def init_kernel_type(self):
@@ -144,6 +146,11 @@ class TestSoftmaxFP16CUDNNOp2(TestSoftmaxFP16CUDNNOp):
         return [2, 3, 4, 5]
 
 
+class TestSoftmaxInference(TestSoftmaxOp):
+    def init_kernel_type(self):
+        self.is_test = True
+
+
 class TestSoftmaxMKLDNNOp(TestSoftmaxOp):
     def init_kernel_type(self):
         self.use_mkldnn = True