diff --git a/cavis-native/cavis-native-lib/build.gradle b/cavis-native/cavis-native-lib/build.gradle
index dcbabf612..759d86205 100644
--- a/cavis-native/cavis-native-lib/build.gradle
+++ b/cavis-native/cavis-native-lib/build.gradle
@@ -251,8 +251,7 @@ chipList.each { String thisChip ->
                             '--chip-extension', avxExtension,
                             '-j', "${host_cores}",
                             // '--helper', 'mkldnn',
-                            '--helper', 'cudnn',
-                            '--tests']
+                            '--helper', 'cudnn']
         } else if (thisChip.equals('cuda') && osdetector.os.startsWith("linux")) { //cuDNN requires CUDA
             it.buildCommand = ['bash', 'buildnativeoperations.sh',
                             '-V',
@@ -273,6 +272,7 @@ chipList.each { String thisChip ->
                             '-j', "${host_cores}",
                             '--helper', 'mkldnn']
         }
+        if(project.hasProperty("nativeTests") it.buildCommand += "--tests+")
     }
 
     //Run the parser on the InfoMap in Nd4j$ChipPresets and listed header files in @Platform
diff --git a/cavis-native/cavis-native-lib/src/main/cpp/blas/CMakeLists.txt b/cavis-native/cavis-native-lib/src/main/cpp/blas/CMakeLists.txt
index e37b2a5ef..d33caee06 100644
--- a/cavis-native/cavis-native-lib/src/main/cpp/blas/CMakeLists.txt
+++ b/cavis-native/cavis-native-lib/src/main/cpp/blas/CMakeLists.txt
@@ -161,13 +161,8 @@ if(HAVE_ARMCOMPUTE)
     file(GLOB_RECURSE CUSTOMOPS_ARMCOMPUTE_SOURCES false ops/declarable/platform/armcompute/*.cpp ops/declarable/platform/armcompute/*.h)
 endif()
 
-if(SD_CUDA)
-    message("Build cublas")
-    if(NOT DEFINED ${CMAKE_CUDA_ARCHITECTURES})
-        set(CMAKE_CUDA_ARCHITECTURES 75)
-    endif()
-    message(STATUS "CUDA architectures set to ${CMAKE_CUDA_ARCHITECTURES}")
 
+if(SD_CUDA)
     find_package(CUDAToolkit)
     enable_language(CUDA)
 
@@ -178,6 +173,8 @@ if(SD_CUDA)
     #Enable features prio C++17
     add_definitions(-D_HAS_AUTO_PTR_ETC=1)
 
+    set(CMAKE_CUDA_RUNTIME_LIBRARY "shared")
+
     #This basically kills instrinsic activated through SD_F16C=true
     #if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
     #    set (CMAKE_CXX_FLAGS "")
@@ -194,7 +191,9 @@ if(SD_CUDA)
             set(EXPM " -D__ND4J_EXPERIMENTAL__=true")
         endif()
 
-
+        if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+          set(CMAKE_CUDA_ARCHITECTURES 75)
+        endif()
 
         # the only difference for debug mode here is host/device debug symbols
         set(CMAKE_CUDA_FLAGS_DEBUG " -G -g")
@@ -217,7 +216,6 @@ if(SD_CUDA)
         endif()
 
         #set(CMAKE_CUDA_FLAGS " ${CMAKE_CUDA_FLAGS} -DCUDA_VERSION_MAJOR=${CUDA_VERSION_MAJOR} ${EXPM} -w --cudart=static --expt-extended-lambda -Xfatbin -compress-all ")
-        set(CMAKE_CUDA_ARCHITECTURES "all-major")
         set(CMAKE_CUDA_RUNTIME_LIBRARY "shared")
         #set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --generate-code \"arch=compute_53,code=[compute_53,sm_53]\" " )
         #set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --generate-code \"arch=compute_61,code=[compute_61,sm_61]\" " )
@@ -240,6 +238,7 @@ if(SD_CUDA)
         message("CUDA_PROPAGATE_HOST_FLAGS = ${CUDA_PROPAGATE_HOST_FLAGS}")
         message("CUDA_ARCH_FLAGS = ${CUDA_ARCH_FLAGS}")
         message("CUDAHOSTCXX = ${CUDAHOSTCXX}")
+        message("CMAKE_CUDA_ARCHITECTURES = ${CMAKE_CUDA_ARCHITECTURES}")
 
         file(GLOB_RECURSE PERF_SOURCES false performance/*.cpp performance/*.h)
         file(GLOB_RECURSE EXCEPTIONS_SOURCES false exceptions/*.cpp exceptions/*.h)