diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec
index 37c2523c9fea85119ae7e6971e4bc5559ea9b8e8..d21cb2697b3368540ea8ca721a7f9e21e04f3d48 100644
--- a/paddle/fluid/API.spec
+++ b/paddle/fluid/API.spec
@@ -378,7 +378,7 @@ paddle.fluid.LoDTensor.__init__ 1. __init__(self: paddle.fluid.core.LoDTensor, a
 paddle.fluid.LoDTensor.has_valid_recursive_sequence_lengths has_valid_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor) -> bool
 paddle.fluid.LoDTensor.lod lod(self: paddle.fluid.core.LoDTensor) -> List[List[int]]
 paddle.fluid.LoDTensor.recursive_sequence_lengths recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor) -> List[List[int]]
-paddle.fluid.LoDTensor.set 1. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CPUPlace) -> None  2. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CPUPlace) -> None  3. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CPUPlace) -> None  4. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CPUPlace) -> None  5. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CPUPlace) -> None  6. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CPUPlace) -> None  7. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CPUPlace) -> None  8. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPlace) -> None  9. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPlace) -> None  10. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPlace) -> None  11. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPlace) -> None  12. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPlace) -> None  13. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPlace) -> None  14. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPlace) -> None  15. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPinnedPlace) -> None  16. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPinnedPlace) -> None  17. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPinnedPlace) -> None  18. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPinnedPlace) -> None  19. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPinnedPlace) -> None  20. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPinnedPlace) -> None  21. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPinnedPlace) -> None
+paddle.fluid.LoDTensor.set 1. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CPUPlace) -> None  2. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CPUPlace) -> None  3. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CPUPlace) -> None  4. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CPUPlace) -> None  5. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CPUPlace) -> None  6. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CPUPlace) -> None  7. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CPUPlace) -> None  8. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CPUPlace) -> None  9. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPlace) -> None  10. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPlace) -> None  11. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPlace) -> None  12. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPlace) -> None  13. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPlace) -> None  14. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPlace) -> None  15. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPlace) -> None  16. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CUDAPlace) -> None  17. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPinnedPlace) -> None  18. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPinnedPlace) -> None  19. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPinnedPlace) -> None  20. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPinnedPlace) -> None  21. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPinnedPlace) -> None  22. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPinnedPlace) -> None  23. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPinnedPlace) -> None  24. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CUDAPinnedPlace) -> None
 paddle.fluid.LoDTensor.set_lod set_lod(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None
 paddle.fluid.LoDTensor.set_recursive_sequence_lengths set_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None
 paddle.fluid.LoDTensor.shape shape(self: paddle.fluid.core.Tensor) -> List[int]
diff --git a/paddle/fluid/framework/data_type.cc b/paddle/fluid/framework/data_type.cc
index 1a9ce746ea840bc088d222cc4e9bc05159d64734..28f3da88fa18021f6b71e458fdb467be86d4dbf0 100644
--- a/paddle/fluid/framework/data_type.cc
+++ b/paddle/fluid/framework/data_type.cc
@@ -64,6 +64,7 @@ static DataTypeMap* InitDataTypeMap() {
   RegType(size_t, proto::VarType::SIZE_T);
   RegType(int16_t, proto::VarType::INT16);
   RegType(uint8_t, proto::VarType::UINT8);
+  RegType(int8_t, proto::VarType::INT8);
 
 #undef RegType
   return retv;
diff --git a/paddle/fluid/framework/data_type.h b/paddle/fluid/framework/data_type.h
index f8c72ffc8964e64a10cff04f322f40b39b2fe263..84691a2059124960a3213802fec0863f8abe6df7 100644
--- a/paddle/fluid/framework/data_type.h
+++ b/paddle/fluid/framework/data_type.h
@@ -54,6 +54,9 @@ inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
     case proto::VarType::INT16:
       visitor.template operator()<int16_t>();
       break;
+    case proto::VarType::INT8:
+      visitor.template operator()<int8_t>();
+      break;
     default:
       PADDLE_THROW("Not supported %d", type);
   }
diff --git a/paddle/fluid/framework/framework.proto b/paddle/fluid/framework/framework.proto
index 2cf14bd371831ab682166f4256d6966b5ab278c8..c6588435819a982166cf2d2368a82b4402fdc2bc 100644
--- a/paddle/fluid/framework/framework.proto
+++ b/paddle/fluid/framework/framework.proto
@@ -107,6 +107,7 @@ message VarType {
     // Tensor<size_t> is used in C++.
     SIZE_T = 19;
     UINT8 = 20;
+    INT8 = 21;
 
     // Other types that may need additional descriptions
     LOD_TENSOR = 7;
diff --git a/paddle/fluid/operators/math/math_function.cc b/paddle/fluid/operators/math/math_function.cc
index c3387be6daa3bd34a6e3410ced23fce5d65f2cf7..9a6e646b28fdec78734eb4e7b98c8acf688b2645 100644
--- a/paddle/fluid/operators/math/math_function.cc
+++ b/paddle/fluid/operators/math/math_function.cc
@@ -41,7 +41,8 @@ template struct SetConstant<platform::CPUDeviceContext, uint8_t>;
   template struct Transpose<platform::CPUDeviceContext, int64_t, RANK>;    \
   template struct Transpose<platform::CPUDeviceContext, bool, RANK>;       \
   template struct Transpose<platform::CPUDeviceContext, int16_t, RANK>;    \
-  template struct Transpose<platform::CPUDeviceContext, uint8_t, RANK>;
+  template struct Transpose<platform::CPUDeviceContext, uint8_t, RANK>;    \
+  template struct Transpose<platform::CPUDeviceContext, int8_t, RANK>;
 
 DEFINE_CPU_TRANS(1);
 DEFINE_CPU_TRANS(2);
diff --git a/paddle/fluid/operators/math/math_function.cu b/paddle/fluid/operators/math/math_function.cu
index d5af718723e8d44da0971ea7756b8c36e771cca2..12d1baa8fb544a8b9684e43204c61ba410d1b295 100644
--- a/paddle/fluid/operators/math/math_function.cu
+++ b/paddle/fluid/operators/math/math_function.cu
@@ -33,10 +33,11 @@ template struct SetConstant<platform::CUDADeviceContext, int>;
 template struct SetConstant<platform::CUDADeviceContext, int64_t>;
 template struct SetConstant<platform::CUDADeviceContext, bool>;
 
-#define DEFINE_GPU_TRANS(RANK)                                          \
-  template struct Transpose<platform::CUDADeviceContext, float, RANK>;  \
-  template struct Transpose<platform::CUDADeviceContext, double, RANK>; \
-  template struct Transpose<platform::CUDADeviceContext, float16, RANK>;
+#define DEFINE_GPU_TRANS(RANK)                                           \
+  template struct Transpose<platform::CUDADeviceContext, float, RANK>;   \
+  template struct Transpose<platform::CUDADeviceContext, double, RANK>;  \
+  template struct Transpose<platform::CUDADeviceContext, float16, RANK>; \
+  template struct Transpose<platform::CUDADeviceContext, int8_t, RANK>;
 
 DEFINE_GPU_TRANS(1);
 DEFINE_GPU_TRANS(2);
diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc
index c2137ec6d7df24251432a4dfb8fffc3d3f77194e..f21f8d23f99c27529b2ed1995c92fd4eee4a5807 100644
--- a/paddle/fluid/pybind/protobuf.cc
+++ b/paddle/fluid/pybind/protobuf.cc
@@ -234,6 +234,7 @@ void BindVarDsec(pybind11::module *m) {
   pybind11::enum_<pd::proto::VarType::Type>(var_desc, "VarType", "")
       .value("BOOL", pd::proto::VarType::BOOL)
       .value("UINT8", pd::proto::VarType::UINT8)
+      .value("INT8", pd::proto::VarType::INT8)
       .value("INT16", pd::proto::VarType::INT16)
       .value("INT32", pd::proto::VarType::INT32)
       .value("INT64", pd::proto::VarType::INT64)
diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index 67734659233515ca8110f4212a2b1553fe4e9d24..5b20b87174e42f4dfdd22214e8f9dd20c7296374 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -130,6 +130,7 @@ PYBIND11_PLUGIN(core) {
       .def("set", PyCPUTensorSetFromArray<bool>)
       .def("set", PyCPUTensorSetFromArray<uint16_t>)
       .def("set", PyCPUTensorSetFromArray<uint8_t>)
+      .def("set", PyCPUTensorSetFromArray<int8_t>)
 #ifdef PADDLE_WITH_CUDA
       .def("set", PyCUDATensorSetFromArray<float>)
       .def("set", PyCUDATensorSetFromArray<int>)
@@ -138,6 +139,7 @@ PYBIND11_PLUGIN(core) {
       .def("set", PyCUDATensorSetFromArray<bool>)
       .def("set", PyCUDATensorSetFromArray<uint16_t>)
       .def("set", PyCUDATensorSetFromArray<uint8_t>)
+      .def("set", PyCUDATensorSetFromArray<int8_t>)
       .def("set", PyCUDAPinnedTensorSetFromArray<float>)
       .def("set", PyCUDAPinnedTensorSetFromArray<int>)
       .def("set", PyCUDAPinnedTensorSetFromArray<double>)
@@ -145,6 +147,7 @@ PYBIND11_PLUGIN(core) {
       .def("set", PyCUDAPinnedTensorSetFromArray<bool>)
       .def("set", PyCUDAPinnedTensorSetFromArray<uint16_t>)
       .def("set", PyCUDAPinnedTensorSetFromArray<uint8_t>)
+      .def("set", PyCUDAPinnedTensorSetFromArray<int8_t>)
 #endif
       .def("shape", [](Tensor &self) { return vectorize(self.dims()); })
       .def("_set_float_element", TensorSetElement<float>)
diff --git a/paddle/fluid/pybind/tensor_py.h b/paddle/fluid/pybind/tensor_py.h
index 3e2ea1ef88b03f5b2576c1cee2b5d26a439943da..51614a6a3dd2f7f830cf533fc365b56a99d3b918 100644
--- a/paddle/fluid/pybind/tensor_py.h
+++ b/paddle/fluid/pybind/tensor_py.h
@@ -97,7 +97,7 @@ struct CastToPyBufferImpl<true, I, ARGS...> {
 inline pybind11::buffer_info CastToPyBuffer(const framework::Tensor &tensor) {
   auto buffer_info =
       details::CastToPyBufferImpl<true, 0, float, int, double, int64_t, bool,
-                                  uint8_t, platform::float16>()(tensor);
+                                  uint8_t, int8_t, platform::float16>()(tensor);
   return buffer_info;
 }
 
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index febb750ee1af26c71e6c1ae1e4c97fb02fb27a04..fbe766336b19719b7a3eac41ad5e877ef3ec4181 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -95,6 +95,8 @@ def convert_np_dtype_to_dtype_(np_dtype):
         return core.VarDesc.VarType.INT16
     elif dtype == np.uint8:
         return core.VarDesc.VarType.UINT8
+    elif dtype == np.int8:
+        return core.VarDesc.VarType.INT8
     else:
         raise ValueError("Not supported numpy dtype %s" % dtype)
 
diff --git a/python/paddle/fluid/tests/unittests/test_tensor.py b/python/paddle/fluid/tests/unittests/test_tensor.py
index e9d0f8a0193c77da33a8cf128dbf8a1c5087782b..1822957c23d0bb1e4821373515d4faef2b76950e 100644
--- a/python/paddle/fluid/tests/unittests/test_tensor.py
+++ b/python/paddle/fluid/tests/unittests/test_tensor.py
@@ -59,6 +59,27 @@ class TestTensor(unittest.TestCase):
         self.assertAlmostEqual(1.0, tensor_array_2[3, 9])
         self.assertAlmostEqual(2.0, tensor_array_2[19, 11])
 
+    def test_int8_tensor(self):
+        scope = core.Scope()
+        var = scope.var("int8_tensor")
+        cpu_tensor = var.get_tensor()
+        tensor_array = numpy.random.randint(
+            -127, high=128, size=[100, 200], dtype=numpy.int8)
+        place = core.CPUPlace()
+        cpu_tensor.set(tensor_array, place)
+        cpu_tensor_array_2 = numpy.array(cpu_tensor)
+        self.assertAlmostEqual(cpu_tensor_array_2.all(), tensor_array.all())
+
+        if core.is_compiled_with_cuda():
+            cuda_tensor = var.get_tensor()
+            tensor_array = numpy.random.randint(
+                -127, high=128, size=[100, 200], dtype=numpy.int8)
+            place = core.CUDAPlace(0)
+            cuda_tensor.set(tensor_array, place)
+            cuda_tensor_array_2 = numpy.array(cuda_tensor)
+            self.assertAlmostEqual(cuda_tensor_array_2.all(),
+                                   tensor_array.all())
+
     def test_int_lod_tensor(self):
         place = core.CPUPlace()
         scope = core.Scope()
diff --git a/python/paddle/fluid/tests/unittests/test_variable.py b/python/paddle/fluid/tests/unittests/test_variable.py
index b0830e130dd9a9037f8dd900a256eea3d05f64b8..4f3c26ca7bdf4d807952b413c8b0dc8b211c06f6 100644
--- a/python/paddle/fluid/tests/unittests/test_variable.py
+++ b/python/paddle/fluid/tests/unittests/test_variable.py
@@ -31,7 +31,8 @@ class TestVariable(unittest.TestCase):
         self.assertEqual(DT.INT16, convert("int16"))
         self.assertEqual(DT.INT64, convert("int64"))
         self.assertEqual(DT.BOOL, convert("bool"))
-        self.assertRaises(ValueError, lambda: convert("int8"))
+        self.assertEqual(DT.INT8, convert("int8"))
+        self.assertEqual(DT.UINT8, convert("uint8"))
 
     def test_var(self):
         b = default_main_program().current_block()