diff --git a/CMakeLists.txt b/CMakeLists.txt
index 0f41982..fd68cd8 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.7)
+cmake_minimum_required(VERSION 3.6)
 project(cuBERT VERSION 0.0.4 LANGUAGES C CXX)
 
 set(CMAKE_CXX_STANDARD 11)
diff --git a/devel-py.Dockerfile b/devel-py.Dockerfile
new file mode 100644
index 0000000..67a7900
--- /dev/null
+++ b/devel-py.Dockerfile
@@ -0,0 +1,21 @@
+FROM quay.io/pypa/manylinux2010_x86_64
+
+WORKDIR /usr/src/cuBERT
+COPY . .
+
+RUN yum -y install \
+        cmake3 \
+        && \
+    yum -y clean all && \
+    rm -rf /var/cache/yum/*
+
+RUN rm -rf build && mkdir build && cd build && \
+    cmake3 -DCMAKE_BUILD_TYPE=Release -DcuBERT_ENABLE_MKL_SUPPORT=ON .. && \
+    make
+
+RUN /opt/python/cp36-cp36m/bin/pip install Cython numpy
+RUN /opt/python/cp27-cp27m/bin/pip install Cython numpy
+
+RUN cd python && \
+    /opt/python/cp36-cp36m/bin/python setup.py bdist_wheel && \
+    /opt/python/cp27-cp27m/bin/python setup.py bdist_wheel
diff --git a/test/cuBERT/BertMTest.cpp b/test/cuBERT/BertMTest.cpp
index 2d1b562..adcfa8d 100644
--- a/test/cuBERT/BertMTest.cpp
+++ b/test/cuBERT/BertMTest.cpp
@@ -1,5 +1,3 @@
-#include "gtest/gtest.h"
-
 #include "common_test.h"
 #include "cuBERT/BertM.h"
 using namespace cuBERT;
diff --git a/test/cuBERT/BertTest.cpp b/test/cuBERT/BertTest.cpp
index 8f4805d..e925bf3 100644
--- a/test/cuBERT/BertTest.cpp
+++ b/test/cuBERT/BertTest.cpp
@@ -1,4 +1,3 @@
-#include "gtest/gtest.h"
 #include <unordered_map>
 
 #include "common_test.h"
diff --git a/test/cuBERT/common_test.cpp b/test/cuBERT/common_test.cpp
index c0849ce..d8d6088 100644
--- a/test/cuBERT/common_test.cpp
+++ b/test/cuBERT/common_test.cpp
@@ -1,5 +1,3 @@
-#include "gtest/gtest.h"
-
 #include "common_test.h"
 
 TEST_F(CommonTest, memcpy) {
diff --git a/test/cuBERT/common_test.h b/test/cuBERT/common_test.h
index c94e27a..3d38ee0 100644
--- a/test/cuBERT/common_test.h
+++ b/test/cuBERT/common_test.h
@@ -1,12 +1,13 @@
 #ifndef CUBERT_COMMON_TEST_H
 #define CUBERT_COMMON_TEST_H
 
-#include "gtest/gtest.h"
 #include <cstdlib>
 #include <cstring>
 
 #include "cuBERT/common.h"
 
+#include "gtest/gtest.h"
+
 class CommonTest : public ::testing::Test {
 protected:
     void SetUp() override {
diff --git a/test/cuBERT/op/DenseTest.cpp b/test/cuBERT/op/DenseTest.cpp
index 5ccb88e..0e51dbc 100644
--- a/test/cuBERT/op/DenseTest.cpp
+++ b/test/cuBERT/op/DenseTest.cpp
@@ -1,5 +1,3 @@
-#include "gtest/gtest.h"
-
 #include "../common_test.h"
 #include "cuBERT/op/Dense.h"
 using namespace cuBERT;
diff --git a/test/cuBERT/op/EmbeddingTest.cpp b/test/cuBERT/op/EmbeddingTest.cpp
index 1460f7c..55bade4 100644
--- a/test/cuBERT/op/EmbeddingTest.cpp
+++ b/test/cuBERT/op/EmbeddingTest.cpp
@@ -1,5 +1,3 @@
-#include "gtest/gtest.h"
-
 #include "../common_test.h"
 #include "cuBERT/op/Embedding.h"
 using namespace cuBERT;
diff --git a/test/cuBERT/op/GELUTest.cpp b/test/cuBERT/op/GELUTest.cpp
index 3028ae4..990f47a 100644
--- a/test/cuBERT/op/GELUTest.cpp
+++ b/test/cuBERT/op/GELUTest.cpp
@@ -1,4 +1,3 @@
-#include "gtest/gtest.h"
 #include <cmath>
 
 #include "../common_test.h"
diff --git a/test/cuBERT/op/LayerNormTest.cpp b/test/cuBERT/op/LayerNormTest.cpp
index af91435..19e2bb6 100644
--- a/test/cuBERT/op/LayerNormTest.cpp
+++ b/test/cuBERT/op/LayerNormTest.cpp
@@ -1,4 +1,3 @@
-#include "gtest/gtest.h"
 #include <cmath>
 
 #include "../common_test.h"
diff --git a/test/cuBERT/op/SoftmaxTest.cpp b/test/cuBERT/op/SoftmaxTest.cpp
index 72a069d..5a8645c 100644
--- a/test/cuBERT/op/SoftmaxTest.cpp
+++ b/test/cuBERT/op/SoftmaxTest.cpp
@@ -1,5 +1,3 @@
-#include "gtest/gtest.h"
-
 #include "../common_test.h"
 #include "cuBERT/op/Softmax.h"
 using namespace cuBERT;
diff --git a/test/cuBERT/op_att/AttentionMaskTest.cpp b/test/cuBERT/op_att/AttentionMaskTest.cpp
index 4c49054..3407273 100644
--- a/test/cuBERT/op_att/AttentionMaskTest.cpp
+++ b/test/cuBERT/op_att/AttentionMaskTest.cpp
@@ -1,5 +1,3 @@
-#include "gtest/gtest.h"
-
 #include "../common_test.h"
 #include "cuBERT/op_att/AttentionMask.h"
 using namespace cuBERT;
diff --git a/test/cuBERT/op_att/AttentionSelfTest.cpp b/test/cuBERT/op_att/AttentionSelfTest.cpp
index a2e8a43..b8c8db5 100644
--- a/test/cuBERT/op_att/AttentionSelfTest.cpp
+++ b/test/cuBERT/op_att/AttentionSelfTest.cpp
@@ -1,4 +1,3 @@
-#include "gtest/gtest.h"
 #include <algorithm>
 
 #include "../common_test.h"
diff --git a/test/cuBERT/op_att/BatchMatMulTest.cpp b/test/cuBERT/op_att/BatchMatMulTest.cpp
index 649e012..b6cb575 100644
--- a/test/cuBERT/op_att/BatchMatMulTest.cpp
+++ b/test/cuBERT/op_att/BatchMatMulTest.cpp
@@ -1,5 +1,3 @@
-#include "gtest/gtest.h"
-
 #include "../common_test.h"
 #include "cuBERT/op_att/BatchMatMul.h"
 using namespace cuBERT;
diff --git a/test/cuBERT/op_att/TransformerTest.cpp b/test/cuBERT/op_att/TransformerTest.cpp
index cf4db48..6ba39e3 100644
--- a/test/cuBERT/op_att/TransformerTest.cpp
+++ b/test/cuBERT/op_att/TransformerTest.cpp
@@ -1,4 +1,3 @@
-#include "gtest/gtest.h"
 #include <algorithm>
 
 #include "../common_test.h"
diff --git a/test/cuBERT/op_bert/BertEmbeddingsTest.cpp b/test/cuBERT/op_bert/BertEmbeddingsTest.cpp
index 0a691ef..3cddaf7 100644
--- a/test/cuBERT/op_bert/BertEmbeddingsTest.cpp
+++ b/test/cuBERT/op_bert/BertEmbeddingsTest.cpp
@@ -1,4 +1,3 @@
-#include "gtest/gtest.h"
 #include <cmath>
 
 #include "../common_test.h"
diff --git a/test/cuBERT/op_bert/BertPoolerTest.cpp b/test/cuBERT/op_bert/BertPoolerTest.cpp
index 57e5965..9b4d4e7 100644
--- a/test/cuBERT/op_bert/BertPoolerTest.cpp
+++ b/test/cuBERT/op_bert/BertPoolerTest.cpp
@@ -1,4 +1,3 @@
-#include "gtest/gtest.h"
 #include <cmath>
 
 #include "../common_test.h"
diff --git a/test/cuBERT/op_out/AdditionalOutputLayerTest.cpp b/test/cuBERT/op_out/AdditionalOutputLayerTest.cpp
index f5ff92b..3fe0361 100644
--- a/test/cuBERT/op_out/AdditionalOutputLayerTest.cpp
+++ b/test/cuBERT/op_out/AdditionalOutputLayerTest.cpp
@@ -1,5 +1,3 @@
-#include "gtest/gtest.h"
-
 #include "../common_test.h"
 #include "cuBERT/op_out/AdditionalOutputLayer.h"
 using namespace cuBERT;