diff --git a/Makefile b/Makefile
index 3d554478b715..e60f57517ad0 100644
--- a/Makefile
+++ b/Makefile
@@ -89,8 +89,6 @@ ifeq ($(USE_DIST_KVSTORE), 1)
 	LDFLAGS += $(PS_LDFLAGS_A)
 endif
 
-include $(MXNET_PLUGINS)
-
 .PHONY: clean all test lint doc clean_all rcpplint rcppexport roxygen
 
 all: lib/libmxnet.a lib/libmxnet.so $(BIN)
@@ -100,6 +98,7 @@ OBJ = $(patsubst %.cc, build/%.o, $(SRC))
 CUSRC = $(wildcard src/*/*.cu)
 CUOBJ = $(patsubst %.cu, build/%_gpu.o, $(CUSRC))
 
+# extra operators
 ifneq ($(EXTRA_OPERATORS),)
 	EXTRA_SRC = $(wildcard $(EXTRA_OPERATORS)/*.cc $(EXTRA_OPERATORS)/*/*.cc)
 	EXTRA_OBJ = $(patsubst $(EXTRA_OPERATORS)/%.cc, $(EXTRA_OPERATORS)/build/%.o, $(EXTRA_SRC))
@@ -113,21 +112,11 @@ else
 endif
 
 # plugin
-ifeq ($(USE_TORCH), 1)
-	CFLAGS += -I$(TORCH_PATH)/install/include -I$(TORCH_PATH)/install/include/TH -I$(TORCH_PATH)/install/include/THC -DMXNET_USE_TORCH=1
-	LDFLAGS += -L$(TORCH_PATH)/install/lib -lluajit -lluaT -lTH -lTHC -L$(TORCH_PATH)/install/lib/lua/5.1 -lpaths -ltorch -lnn
-	ifeq ($(USE_CUDA), 1)
-		LDFLAGS += -lcutorch -lcunn
-	endif
-
-	TORCH_SRC = $(wildcard plugin/torch/*.cc)
-	PLUGIN_OBJ += $(patsubst %.cc, build/%.o, $(TORCH_SRC))
-	TORCH_CUSRC = $(wildcard plugin/torch/*.cu)
-	PLUGIN_CUOBJ += $(patsubst %.cu, build/%_gpu.o, $(TORCH_CUSRC))
-else
-	CFLAGS += -DMXNET_USE_TORCH=0
-endif
+PLUGIN_OBJ =
+PLUGIN_CUOBJ =
+include $(MXNET_PLUGINS)
 
+# all dep
 LIB_DEP += $(DMLC_CORE)/libdmlc.a
 ALL_DEP = $(OBJ) $(EXTRA_OBJ) $(PLUGIN_OBJ) $(LIB_DEP)
 ifeq ($(USE_CUDA), 1)
@@ -158,10 +147,11 @@ build/plugin/%.o: plugin/%.cc
 	$(CXX) -std=c++0x $(CFLAGS) -MM -MT build/plugin/$*.o $< >build/plugin/$*.d
 	$(CXX) -std=c++0x -c $(CFLAGS) -c $< -o $@
 
-# A nvcc bug cause this to generate "generic/xxx.h" dependencies from torch headers.
-# $(NVCC) $(NVCCFLAGS) -Xcompiler "$(CFLAGS)" -M -MT build/plugin/$*_gpu.o $< >build/plugin/$*_gpu.d
+# A nvcc bug cause it to generate "generic/xxx.h" dependencies from torch headers.
+# Use CXX to generate dependency instead.
 build/plugin/%_gpu.o: plugin/%.cu
 	@mkdir -p $(@D)
+	$(CXX) -std=c++0x $(CFLAGS) -MM -MT build/plugin/$*_gpu.o $< >build/plugin/$*_gpu.d
 	$(NVCC) -c -o $@ $(NVCCFLAGS) -Xcompiler "$(CFLAGS)" $<
 
 $(EXTRA_OPERATORS)/build/%.o: $(EXTRA_OPERATORS)/%.cc
diff --git a/make/config.mk b/make/config.mk
index 484f60066a72..7bba2101de1f 100644
--- a/make/config.mk
+++ b/make/config.mk
@@ -112,8 +112,8 @@ EXTRA_OPERATORS =
 #----------------------------
 
 # whether to use torch integration. This requires installing torch.
-USE_TORCH = 0
-TORCH_PATH = $(HOME)/torch
+# TORCH_PATH = $(HOME)/torch
+# MXNET_PLUGINS += plugin/torch/torch.mk
 
 # whether to use sframe integration. This requires build sframe
 # git@github.com:dato-code/SFrame.git
diff --git a/make/osx.mk b/make/osx.mk
index 3995049e60b1..5a6f56cc8ca2 100644
--- a/make/osx.mk
+++ b/make/osx.mk
@@ -99,5 +99,5 @@ EXTRA_OPERATORS =
 #----------------------------
 
 # whether to use torch integration. This requires installing torch.
-USE_TORCH = 0
-TORCH_PATH = $(HOME)/torch
+# TORCH_PATH = $(HOME)/torch
+# MXNET_PLUGINS += plugin/torch/torch.mk
diff --git a/plugin/torch/torch.mk b/plugin/torch/torch.mk
new file mode 100644
index 000000000000..5eda7e0a015a
--- /dev/null
+++ b/plugin/torch/torch.mk
@@ -0,0 +1,10 @@
+CFLAGS += -I$(TORCH_PATH)/install/include -DMXNET_USE_TORCH=1
+LDFLAGS += -L$(TORCH_PATH)/install/lib -lluajit -lluaT -lTH -lTHC -L$(TORCH_PATH)/install/lib/lua/5.1 -lpaths -ltorch -lnn
+ifeq ($(USE_CUDA), 1)
+	LDFLAGS += -lcutorch -lcunn
+endif
+
+TORCH_SRC = $(wildcard plugin/torch/*.cc)
+PLUGIN_OBJ += $(patsubst %.cc, build/%.o, $(TORCH_SRC))
+TORCH_CUSRC = $(wildcard plugin/torch/*.cu)
+PLUGIN_CUOBJ += $(patsubst %.cu, build/%_gpu.o, $(TORCH_CUSRC))
\ No newline at end of file
diff --git a/plugin/torch/torch_base.h b/plugin/torch/torch_base.h
index ecc6d67e6ed6..74ede4d4c1a1 100644
--- a/plugin/torch/torch_base.h
+++ b/plugin/torch/torch_base.h
@@ -7,23 +7,24 @@
 #ifndef PLUGIN_TORCH_TORCH_BASE_H_
 #define PLUGIN_TORCH_TORCH_BASE_H_
 #include <mxnet/base.h>
-#include <vector>
 
 extern "C" {
 #include <lua.h>
 #include <luaT.h>
 #include <lualib.h>
-#include <THStorage.h>
-#include <THTensor.h>
+#include <TH/THStorage.h>
+#include <TH/THTensor.h>
 }
 
 #if MXNET_USE_CUDA
 extern "C" {
-#include <THCStorage.h>
-#include <THCTensor.h>
+#include <THC/THCStorage.h>
+#include <THC/THCTensor.h>
 }
 #endif  // MXNET_USE_CUDA
 
+#include <vector>
+
 namespace mxnet {
 
 class TorchState {