diff --git a/src/configure b/src/configure
index c4a1445efbd..ebd9e8f62c3 100755
--- a/src/configure
+++ b/src/configure
@@ -462,7 +462,8 @@ function configure_cuda {
     elif [ "`uname -m`" == "ppc64le" ]; then
       cat makefiles/cuda_64bit.mk >> kaldi.mk
     else
-      cat makefiles/cuda_32bit.mk >> kaldi.mk
+      echo "CUDA will not be used! CUDA is not supported with 32-bit builds."
+      exit 1;
     fi
 
   else
diff --git a/src/makefiles/cuda_32bit.mk b/src/makefiles/cuda_32bit.mk
deleted file mode 100644
index 9dcc965d7b6..00000000000
--- a/src/makefiles/cuda_32bit.mk
+++ /dev/null
@@ -1,15 +0,0 @@
-ifndef DOUBLE_PRECISION
-$(error DOUBLE_PRECISION not defined.)
-endif
-ifndef CUDATKDIR
-$(error CUDATKDIR not defined.)
-endif
-
-CUDA_INCLUDE= -I$(CUDATKDIR)/include -I$(CUBROOT)
-CUDA_FLAGS = -Xcompiler "-fPIC -pthread -isystem $(OPENFSTINC)" --verbose --machine 32 -DHAVE_CUDA \
-             -ccbin $(CXX) -DKALDI_DOUBLEPRECISION=$(DOUBLE_PRECISION) \
-             -std=c++11 -DCUDA_API_PER_THREAD_DEFAULT_STREAM
-
-CXXFLAGS += -DHAVE_CUDA -I$(CUDATKDIR)/include
-CUDA_LDFLAGS += -L$(CUDATKDIR)/lib -Wl,-rpath,$(CUDATKDIR)/lib
-CUDA_LDLIBS += -lcublas -lcusparse -lcudart -lcurand -lnvToolsExt #LDLIBS : The libs are loaded later than static libs in implicit rule
diff --git a/src/makefiles/cuda_64bit.mk b/src/makefiles/cuda_64bit.mk
index 72314d45622..dbe0b6d9416 100644
--- a/src/makefiles/cuda_64bit.mk
+++ b/src/makefiles/cuda_64bit.mk
@@ -5,11 +5,13 @@ ifndef CUDATKDIR
 $(error CUDATKDIR not defined.)
 endif
 
+CXXFLAGS += -DHAVE_CUDA -I$(CUDATKDIR)/include -fPIC -pthread -isystem $(OPENFSTINC) -rdynamic --verbose
+
 CUDA_INCLUDE= -I$(CUDATKDIR)/include -I$(CUBROOT)
-CUDA_FLAGS = -Xcompiler "-fPIC -pthread -isystem $(OPENFSTINC)" --verbose --machine 64 -DHAVE_CUDA \
+CUDA_FLAGS = --machine 64 -DHAVE_CUDA \
              -ccbin $(CXX) -DKALDI_DOUBLEPRECISION=$(DOUBLE_PRECISION) \
-             -std=c++11 -DCUDA_API_PER_THREAD_DEFAULT_STREAM
+             -std=c++11 -DCUDA_API_PER_THREAD_DEFAULT_STREAM  -lineinfo \
+						 -Xcompiler "$(CXXFLAGS)"
 
-CXXFLAGS += -DHAVE_CUDA -I$(CUDATKDIR)/include
 CUDA_LDFLAGS += -L$(CUDATKDIR)/lib64 -Wl,-rpath,$(CUDATKDIR)/lib64
 CUDA_LDLIBS += -lcublas -lcusparse -lcudart -lcurand -lnvToolsExt #LDLIBS : The libs are loaded later than static libs in implicit rule
diff --git a/src/makefiles/default_rules.mk b/src/makefiles/default_rules.mk
index ee0f3c2e90b..956e0eedc33 100644
--- a/src/makefiles/default_rules.mk
+++ b/src/makefiles/default_rules.mk
@@ -114,8 +114,30 @@ valgrind: .valgrind
 	rm valgrind.out
 	touch .valgrind
 
+
+#buid up dependency commands
+CC_SRCS=$(wildcard *.cc)
+#check if files exist to run dependency commands on
+ifneq ($(CC_SRCS),)										
+CC_DEP_COMMAND=$(CXX) -M $(CXXFLAGS) $(CC_SRCS)
+endif
+
+ifeq ($(CUDA), true)
+CUDA_SRCS=$(wildcard *.cu)
+#check if files exist to run dependency commands on
+ifneq ($(CUDA_SRCS),)
+NVCC_DEP_COMMAND = $(CUDATKDIR)/bin/nvcc -M $(CUDA_FLAGS) $(CUDA_INCLUDE) $(CUDA_SRCS)
+endif
+endif
+
 depend:
-	-$(CXX) -M $(CXXFLAGS) *.cc > .depend.mk
+	rm -f .depend.mk
+ifneq ($(CC_DEP_COMMAND),)
+	$(CC_DEP_COMMAND) >> .depend.mk
+endif
+ifneq ($(NVCC_DEP_COMMAND),)
+	$(NVCC_DEP_COMMAND) >> .depend.mk
+endif
 
 # removing automatic making of "depend" as it's quite slow.
 #.depend.mk: depend