diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9c2075deec6344db3f56effc01ff4f62460ae20a..1d407edb35425aca609802df4672ecf6104bddbd 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -261,6 +261,7 @@ if(MGE_WITH_CUDA)
     set(MGE_CUDA_LIBS "${MGE_CUDA_LIBS}")
 endif()
 
+
 find_program(CCACHE_BIN ccache)
 if(CCACHE_BIN)
     set(CMAKE_CXX_COMPILER_LAUNCHER ${CCACHE_BIN})
diff --git a/dnn/src/CMakeLists.txt b/dnn/src/CMakeLists.txt
index 2defb17e39c4e9eddaed2a0c7f6c2f0f6f3c6cc6..f0cbbe7994e9f54d06e613e6535f49e128695021 100644
--- a/dnn/src/CMakeLists.txt
+++ b/dnn/src/CMakeLists.txt
@@ -56,4 +56,3 @@ target_link_libraries(megdnn ${MGE_BLAS_LIBS})
 if(CMAKE_THREAD_LIBS_INIT)
     target_link_libraries(megdnn Threads::Threads)
 endif()
-
diff --git a/dnn/src/common/megcore/common/device_context.cpp b/dnn/src/common/megcore/common/device_context.cpp
index f66da9a79affcde2858b1df81600d7e78b51e5e7..ee4114c8141b4f4aa96e0aad17660e37b4a32dcd 100644
--- a/dnn/src/common/megcore/common/device_context.cpp
+++ b/dnn/src/common/megcore/common/device_context.cpp
@@ -16,7 +16,6 @@
 #include "src/cuda/megcore/cuda_device_context.hpp"
 #endif
 
-
 using namespace megcore;
 using namespace megdnn;
 
diff --git a/dnn/test/CMakeLists.txt b/dnn/test/CMakeLists.txt
index ddbba508aeca3b6c6b804eaf906f1250d3830f19..c1f40a4183700fdef779917e46508b95195193ec 100644
--- a/dnn/test/CMakeLists.txt
+++ b/dnn/test/CMakeLists.txt
@@ -26,6 +26,7 @@ if(MGE_WITH_CUDA)
 endif()
 
 
+
 add_executable(megdnn_test ${SOURCES})
 target_link_libraries(megdnn_test gtest)
 target_link_libraries(megdnn_test megdnn)
diff --git a/python_module/test/run.sh b/python_module/test/run.sh
index 4d821340471966a931b1b3345fb33e3b07badec3..c13066003372d719aeeeb804259ee8cf62869be6 100755
--- a/python_module/test/run.sh
+++ b/python_module/test/run.sh
@@ -9,5 +9,6 @@ pushd $(dirname "${BASH_SOURCE[0]}")/.. >/dev/null
         --ignore test/unit/data \
         --ignore test/integration/manual \
         --ignore megengine/module/pytorch \
+        --ignore test/unit/module/test_external.py \
         megengine test
 popd >/dev/null
diff --git a/python_module/test/unit/module/.gitattributes b/python_module/test/unit/module/.gitattributes
new file mode 100644
index 0000000000000000000000000000000000000000..816f72d606d559356d757eea2817f579b9ebbeef
--- /dev/null
+++ b/python_module/test/unit/module/.gitattributes
@@ -0,0 +1 @@
+*.mlu binary
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 7b7f05b79695d28853c97208acfbbf265514327b..221b071a33e22921f22c1ae7905d8d73c2315a3a 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -31,6 +31,7 @@ if(MGE_WITH_CUDA AND MGE_WITH_TRT)
     list(APPEND SOURCES ${SOURCES_})
 endif()
 
+
 set(MGB_DEF ${MGB_DEF} PARENT_SCOPE)
 add_library(megbrain STATIC EXCLUDE_FROM_ALL ${SOURCES})
 target_link_libraries(megbrain mgb_opr_param_defs)
diff --git a/src/core/impl/comp_node_env.cpp b/src/core/impl/comp_node_env.cpp
index 993a83c2aca523ff06d2292e807bbff4df7889f0..9c8bfb899e46c2c266439c62e782f471608de012 100644
--- a/src/core/impl/comp_node_env.cpp
+++ b/src/core/impl/comp_node_env.cpp
@@ -22,7 +22,6 @@
 #endif
 #endif
 
-
 using namespace mgb;
 
 /* =================== MegDNNHandle =================== */
diff --git a/src/core/include/megbrain/comp_node/alloc.h b/src/core/include/megbrain/comp_node/alloc.h
index 0c3c66e5d40445872572b32a2f54a908d9745adb..9d7e51f064331b486f5b13568cbeb43e4dc9678c 100644
--- a/src/core/include/megbrain/comp_node/alloc.h
+++ b/src/core/include/megbrain/comp_node/alloc.h
@@ -200,7 +200,6 @@ class DevMemAlloc: virtual public MemAllocBase {
 #endif
 
 
-
         virtual ~DevMemAlloc() = default;
 
         /*!
diff --git a/src/core/include/megbrain/comp_node_env.h b/src/core/include/megbrain/comp_node_env.h
index 1c0b37cb65731f8206c287491086de19f9d58932..fd0e846c7c560a38f140ece0b4fdd1e762a6c2b1 100644
--- a/src/core/include/megbrain/comp_node_env.h
+++ b/src/core/include/megbrain/comp_node_env.h
@@ -41,7 +41,7 @@
         }                                                               \
     } while (0)
 
-#endif //MGB_ENABLE_LOGGING
+#endif  // MGB_ENABLE_LOGGING
 
 #endif
 
diff --git a/src/core/include/megbrain_build_config.h.in b/src/core/include/megbrain_build_config.h.in
index aa5612b9b0cabcf4e7018ee99a3e9823d4eefeaf..544848423acd48ae200f1332198d7d5a57d3d633 100644
--- a/src/core/include/megbrain_build_config.h.in
+++ b/src/core/include/megbrain_build_config.h.in
@@ -97,6 +97,10 @@
 #endif
 
 
+#ifndef MGB_CAMBRICON
+#define MGB_CAMBRICON 0
+#endif
+
 // whether to enable TensorRT support
 #ifndef MGB_ENABLE_TENSOR_RT
 #define MGB_ENABLE_TENSOR_RT    MGB_CUDA
diff --git a/src/core/test/comp_node.cpp b/src/core/test/comp_node.cpp
index b4559a62fc4243d0a3e3ac351f6adb7f98fce277..465ca6c8fa4ed14701cebfbac4231b67cf4d7fd3 100644
--- a/src/core/test/comp_node.cpp
+++ b/src/core/test/comp_node.cpp
@@ -49,7 +49,8 @@ TEST(TestCompNode, Parse) {
     ASSERT_EQ(L::parse("cpu2:23"), make_lc(D::CPU, 2, 23));
     ASSERT_EQ(L::parse("cpu21:23"), make_lc(D::CPU, 21, 23));
 
-    ASSERT_EQ(L::parse("xpu"), make_lc(D::UNSPEC, -1, 0)); 
+    
+    ASSERT_EQ(L::parse("xpu"), make_lc(D::UNSPEC, -1, 0));
     ASSERT_EQ(L::parse("xpux"), make_lc(D::UNSPEC, -1, 0));
     ASSERT_EQ(L::parse("xpu23"), make_lc(D::UNSPEC, 23, 0));
     ASSERT_EQ(L::parse("xpu23:1"), make_lc(D::UNSPEC, 23, 1));
@@ -70,6 +71,7 @@ TEST(TestCompNode, Parse) {
     ASSERT_THROW(L::parse("cpu2:23x"), MegBrainError);
     ASSERT_THROW(L::parse("heaxgon0"), MegBrainError);
     ASSERT_THROW(L::parse("rcom0"), MegBrainError);
+    ASSERT_THROW(L::parse("cmabricon0"), MegBrainError);
 }
 
 TEST(TestCompNode, SetDefaultDev) {
@@ -546,6 +548,7 @@ TEST(TestCompNode, MultipleLoad) {
     }
 }
 
+
 namespace {
 class CompNodeDepedentObjectInst final : public CompNodeDepedentObject {
     int *m_dst, *m_timer;
diff --git a/src/core/test/mem_alloc.cpp b/src/core/test/mem_alloc.cpp
index ba8ad6714a3cec2fd9001bedb31e8192c8c92013..83b07ae39a19a0ebe0b724c25ac1c5e923cbb520 100644
--- a/src/core/test/mem_alloc.cpp
+++ b/src/core/test/mem_alloc.cpp
@@ -464,6 +464,7 @@ public:
     }
     void raw_dev_free(void* ptr) override { MGB_CUDA_CHECK(cudaFree(ptr)); }
 };
+#endif
 
 using Callback = std::function<void()>;
 void test_free_mem(CompNode cn0, CompNode cn1, DevicePolicy* policy,
@@ -529,7 +530,7 @@ void test_gather_other(CompNode cn0, CompNode cn1) {
     opr::Sleep::sleep(cn1, 0.7);
     func->execute();
 }
-#endif
+
 }  // namespace
 
 #if MGB_CUDA
@@ -562,4 +563,5 @@ TEST(TestCudaMemAlloc, FreeMem) {
 }
 #endif  // MGB_CUDA
 
+
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}
diff --git a/test/src/helper.cpp b/test/src/helper.cpp
index d88429cdb4b2fd55023b802cf4ca37f04fc58953..d8acb57487967bf1d383df133d681950b676b9e3 100644
--- a/test/src/helper.cpp
+++ b/test/src/helper.cpp
@@ -70,8 +70,8 @@ dtype, RandomDistribution::GAUSSIAN>::operator ()(
     auto ptr = ret->ptr<ctype>();
     auto mean = m_mean, std = m_std;
     for (size_t i = 0, it = shape.total_nr_elems(); i < it; i += 2) {
-        ctype u1 = (m_rng() + 1.0) / (m_rng.max() + 1.0),
-              u2 = (m_rng() + 1.0) / (m_rng.max() + 1.0),
+        ctype u1 = ctype((m_rng() + 1.0) / (m_rng.max() + 1.0)),
+              u2 = ctype((m_rng() + 1.0) / (m_rng.max() + 1.0)),
               r = ctype(std * std::sqrt(-2 * std::log(u1))),
               theta = ctype(2 * M_PI * u2),
               z0 = ctype(r * std::cos(theta) + mean),
@@ -104,6 +104,8 @@ namespace mgb {
         dtype::Float32, RandomDistribution::GAUSSIAN>;
     template class HostTensorGenerator<
         dtype::Float32, RandomDistribution::UNIFORM>;
+    template class HostTensorGenerator<
+        dtype::Float16, RandomDistribution::GAUSSIAN>;
     template class HostTensorGenerator<
         dtype::Int8, RandomDistribution::UNIFORM>;
     template class HostTensorGenerator<
diff --git a/test/src/include/megbrain/test/helper.h b/test/src/include/megbrain/test/helper.h
index d8b5fcfb4868021d2918384f6b862a4d998a18ea..eaea3e2efb22254a19311ca18adf774c77d09289 100644
--- a/test/src/include/megbrain/test/helper.h
+++ b/test/src/include/megbrain/test/helper.h
@@ -400,6 +400,9 @@ bool check_gpu_available(size_t num);
 //! check whether given number of AMD GPUs is available
 bool check_amd_gpu_available(size_t num);
 
+//! check whether given number of cambricon devices is available
+bool check_cambricon_device_available(size_t num);
+
 //! check current capability >= major.minor
 bool check_compute_capability(int major, int minor);
 
@@ -436,6 +439,7 @@ public:
         return; \
 } while(0)
 
+
 #if MGB_HAVE_THREAD
 #define REQUIRE_THREAD()
 #else