diff --git a/inc/external/acl/acl_rt.h b/inc/external/acl/acl_rt.h
index 7d6455b164380208c0570449b1bffddf56495440..38dfc2528f93c8b891d759653c286aa6ef959b68 100644
--- a/inc/external/acl/acl_rt.h
+++ b/inc/external/acl/acl_rt.h
@@ -87,6 +87,12 @@ typedef enum aclrtGroupAttr {
     ACL_GROUP_GROUPID_INT
 } aclrtGroupAttr;
 
+typedef enum aclrtFloatOverflowMode {
+    ACL_RT_OVERFLOW_MODE_SATURATION = 0,
+    ACL_RT_OVERFLOW_MODE_INFNAN,
+    ACL_RT_OVERFLOW_MODE_UNDEF,
+} aclrtFloatOverflowMode;
+
 typedef struct tagRtGroupInfo aclrtGroupInfo;
 
 typedef struct rtExceptionInfo aclrtExceptionInfo;
@@ -1028,6 +1034,48 @@ ACL_FUNC_VISIBILITY aclError aclrtGetMemInfo(aclrtMemAttr attr, size_t *free, si
  */
 ACL_FUNC_VISIBILITY aclError aclrtSetOpWaitTimeout(uint32_t timeout);
 
+/**
+ * @ingroup AscendCL
+ * @brief enable or disable overflow switch on some stream
+ * @param stream [IN]   set overflow switch on this stream
+ * @param flag [IN]  0 : disable 1 : enable
+ *
+ * @retval ACL_SUCCESS The function is successfully executed.
+ * @retval OtherValues Failure
+ */
+ACL_FUNC_VISIBILITY aclError aclrtSetStreamOverflowSwitch(aclrtStream stream, uint32_t flag);
+
+/**
+ * @ingroup AscendCL
+ * @brief get overflow switch on some stream
+ * @param stream [IN]   get overflow switch on this stream
+ * @param flag [OUT]  current overflow switch, 0 : disable others : enable
+ *
+ * @retval ACL_SUCCESS The function is successfully executed.
+ * @retval OtherValues Failure
+ */
+ACL_FUNC_VISIBILITY aclError aclrtGetStreamOverflowSwitch(aclrtStream stream, uint32_t *flag);
+
+/**
+ * @ingroup AscendCL
+ * @brief set saturation mode
+ * @param mode [IN]   target saturation mode
+ *
+ * @retval ACL_SUCCESS The function is successfully executed.
+ * @retval OtherValues Failure
+ */
+ACL_FUNC_VISIBILITY aclError aclrtSetDeviceSatMode(aclrtFloatOverflowMode mode);
+
+/**
+ * @ingroup AscendCL
+ * @brief get saturation mode
+ * @param mode [OUT]   get saturation mode
+ *
+ * @retval ACL_SUCCESS The function is successfully executed.
+ * @retval OtherValues Failure
+ */
+ACL_FUNC_VISIBILITY aclError aclrtGetDeviceSatMode(aclrtFloatOverflowMode *mode);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/tf_adapter/python/npu_bridge/estimator/npu/npu_plugin.py b/tf_adapter/python/npu_bridge/estimator/npu/npu_plugin.py
index 63ad87065c42ab585cad7260b075e72dab58e7a7..2c5d9a107c0391ea3694da205016599a76338349 100644
--- a/tf_adapter/python/npu_bridge/estimator/npu/npu_plugin.py
+++ b/tf_adapter/python/npu_bridge/estimator/npu/npu_plugin.py
@@ -226,3 +226,7 @@ def get_rdma_cache(data_type, shape, name="rdma_w"):
     with npu_scope.npu_mem_type_scope():
         return variable_scope.get_variable(name=name, shape=shape, dtype=data_type,
                                            initializer=init_ops.zeros_initializer())
+
+
+def set_device_sat_mode(mode):
+    tf_adapter.SetDeviceSatMode(mode)
diff --git a/tf_adapter/swig/ge_plugin.i b/tf_adapter/swig/ge_plugin.i
index bb666ab62d200402c694432956660a86b226bc43..61b5ade1114b723e73a3f7b03550005acb0e29f4 100644
--- a/tf_adapter/swig/ge_plugin.i
+++ b/tf_adapter/swig/ge_plugin.i
@@ -51,6 +51,10 @@ extern int32_t RdmaInitAndRegister(const std::vector<ge::HostVarInfo> &var_info,
 extern int32_t GetVarAddrAndSize(const std::string &var_name, uint64_t &base_addr, uint64_t &var_size);
 
 extern int32_t MallocSharedMem(const ge::TensorInfo &tensor_info, uint64_t &dev_addr, uint64_t &memory_size);
+
+extern int32_t SetDeviceSatMode(uint32_t mode);
+
+extern int32_t GetDeviceSatMode();
 %}
 
 %template(var_info_vec) std::vector<ge::HostVarInfo>;
@@ -120,3 +124,7 @@ extern int32_t RdmaInitAndRegister(const std::vector<ge::HostVarInfo> &var_info,
 extern int32_t GetVarAddrAndSize(const std::string &var_name, uint64_t &base_addr, uint64_t &var_size);
 
 extern int32_t MallocSharedMem(const ge::TensorInfo &tensor_info, uint64_t &dev_addr, uint64_t &memory_size);
+
+extern int32_t SetDeviceSatMode(uint32_t mode);
+
+extern int32_t GetDeviceSatMode();
diff --git a/tf_adapter/tests/depends/ascendcl/src/ascendcl_stub.cc b/tf_adapter/tests/depends/ascendcl/src/ascendcl_stub.cc
index 277d68aecd22a777b1bdc04cc54e4eb50893d5cc..197aa4c3343e3617a1cda345e05c3c0b5e835c84 100644
--- a/tf_adapter/tests/depends/ascendcl/src/ascendcl_stub.cc
+++ b/tf_adapter/tests/depends/ascendcl/src/ascendcl_stub.cc
@@ -24,6 +24,8 @@
 #include "tf_adapter/common/adapter_logger.h"
 
 namespace {
+    constexpr uint32_t kDeviceSatModeLimit = 2U;
+    std::uint32_t deviceSatMode = 2U;
     std::mutex aclChannleMutex;
     std::map<std::string, acltdtChannelHandle *> aclChannleMap;
     std::map<std::string, aclDataType> aclDataTypeStrMap =
@@ -592,3 +594,20 @@ aclError aclmdlExecuteAsync(uint32_t modelId, const aclmdlDataset *inputs, aclmd
   }
   return ACL_SUCCESS;
 }
+
+aclError aclrtSetDeviceSatMode(aclrtFloatOverflowMode mode) {
+  if (mode != ACL_RT_OVERFLOW_MODE_SATURATION && mode != ACL_RT_OVERFLOW_MODE_INFNAN) {
+    deviceSatMode = 2U;
+    return ACL_ERROR_INVALID_PARAM;
+  }
+  deviceSatMode = static_cast<uint32_t>(mode);
+  return ACL_SUCCESS;
+}
+
+aclError aclrtGetDeviceSatMode(aclrtFloatOverflowMode *mode) {
+  if (deviceSatMode >= kDeviceSatModeLimit) {
+    return ACL_ERROR_FAILURE;
+  }
+  *mode = static_cast<aclrtFloatOverflowMode>(deviceSatMode);
+  return ACL_SUCCESS;
+}
diff --git a/tf_adapter/tests/st/util/testcase/ge_plugin_test.cc b/tf_adapter/tests/st/util/testcase/ge_plugin_test.cc
index 754faf743f5a7cd95f592f1735e046b006c0b6d4..b53da180075ee08ffff95965643f3285439798e5 100644
--- a/tf_adapter/tests/st/util/testcase/ge_plugin_test.cc
+++ b/tf_adapter/tests/st/util/testcase/ge_plugin_test.cc
@@ -127,6 +127,16 @@ TEST_F(GePluginTest, MallocSharedMemOKTest) {
   int32_t ret = MallocSharedMem(tensor_info, dev_addr, memory_size);
   EXPECT_EQ(ret, 0);
 }
+TEST_F(GePluginTest, SetDeviceSatModeTest) {
+  uint32_t mode = 1U;
+  int32_t ret = SetDeviceSatMode(mode);
+  EXPECT_EQ(ret, 0);
+  EXPECT_EQ(GetDeviceSatMode(), 1);
+  mode = 2U;
+  ret = SetDeviceSatMode(mode);
+  EXPECT_EQ(ret, -1);
+  EXPECT_EQ(GetDeviceSatMode(), -1);
+}
 TEST_F(GePluginTest, NpuCloseTest) {
   std::map<std::string, std::string> init_options;
   init_options["ge.jobType"] = "1";
@@ -159,4 +169,4 @@ TEST_F(GePluginTest, RdmaInitAndRegisterOKTest) {
   EXPECT_EQ(ret, 0);
 }
 }
-} // end tensorflow
\ No newline at end of file
+} // end tensorflow
diff --git a/tf_adapter/tests/ut/util/testcase/ge_plugin_test.cc b/tf_adapter/tests/ut/util/testcase/ge_plugin_test.cc
index bb0bf3c4ee5e7079d7f5c8ba5f69520d4027df43..7b1bbba005ac3a7f7ffcb39eeba1c904a1ae1b06 100644
--- a/tf_adapter/tests/ut/util/testcase/ge_plugin_test.cc
+++ b/tf_adapter/tests/ut/util/testcase/ge_plugin_test.cc
@@ -122,6 +122,16 @@ TEST_F(GePluginTest, MallocSharedMemOKTest) {
   int32_t ret = MallocSharedMem(tensor_info, dev_addr, memory_size);
   EXPECT_EQ(ret, 0);
 }
+TEST_F(GePluginTest, SetDeviceSatModeTest) {
+  uint64_t mode = 1U;
+  int32_t ret = SetDeviceSatMode(mode);
+  EXPECT_EQ(ret, 0);
+  EXPECT_EQ(GetDeviceSatMode(), 1);
+  mode = 2U;
+  ret = SetDeviceSatMode(mode);
+  EXPECT_EQ(ret, -1);
+  EXPECT_EQ(GetDeviceSatMode(), -1);
+}
 TEST_F(GePluginTest, NpuCloseTest) {
   std::map<std::string, std::string> init_options;
   init_options["ge.jobType"] = "1";
@@ -155,4 +165,4 @@ TEST_F(GePluginTest, RdmaInitAndRegisterOKTest) {
 }
 
 }
-} // end tensorflow
\ No newline at end of file
+} // end tensorflow
diff --git a/tf_adapter/util/ge_plugin.cc b/tf_adapter/util/ge_plugin.cc
index cd8c8e5d7654f0937bcf9d4de8ef357519ec1a10..86532ef359f59b358da66dd5c56a6bcf457e4579 100644
--- a/tf_adapter/util/ge_plugin.cc
+++ b/tf_adapter/util/ge_plugin.cc
@@ -27,6 +27,7 @@
 #include "tf_adapter/common/adapter_logger.h"
 #include "tf_adapter/common/common.h"
 #include "tf_adapter/util/npu_attrs.h"
+#include "acl/acl_rt.h"
 #include "tf_adapter/util/npu_plugin.h"
 #include "aoe_tuning_api.h"
 using AoeFinalizeFunc = Aoe::AoeStatus (*)();
@@ -478,4 +479,26 @@ int32_t MallocSharedMem(const ge::TensorInfo &tensor_info, uint64_t &dev_addr, u
   return 0;
 }
 
-std::atomic_int GePlugin::graph_counter_ = {0};
\ No newline at end of file
+int32_t SetDeviceSatMode(uint32_t mode) {
+  aclError ret = aclrtSetDeviceSatMode(aclrtFloatOverflowMode(mode));
+  if (ret != ACL_SUCCESS) {
+    ADP_LOG(ERROR) << "[GePlugin] set device sat mode failed, ret : " << ToString(ret);
+    LOG(ERROR) << "[GePlugin] set device sat mode failed, ret : " << ToString(ret);
+    return -1;
+  }
+  ADP_LOG(INFO) << "[GePlugin] set device sat mode success.";
+  return 0;
+}
+
+int32_t GetDeviceSatMode() {
+  aclrtFloatOverflowMode floatOverflowMode = ACL_RT_OVERFLOW_MODE_UNDEF;
+  aclError ret = aclrtGetDeviceSatMode(&floatOverflowMode);
+  if (ret != ACL_SUCCESS) {
+    ADP_LOG(ERROR) << "[GePlugin] get device sat mode failed, ret : " << ToString(ret);
+    LOG(ERROR) << "[GePlugin] get device sat mode failed, ret : " << ToString(ret);
+    return -1;
+  }
+  ADP_LOG(INFO) << "[GePlugin] get device sat mode success.";
+  return static_cast<int32_t>(floatOverflowMode);
+}
+std::atomic_int GePlugin::graph_counter_ = {0};
diff --git a/tf_adapter/util/npu_plugin.h b/tf_adapter/util/npu_plugin.h
index c2a4f2fbd42ff737728b5fd378f844e97b6496d1..a76ea92b9a94b7d096130388515d58915d710186 100644
--- a/tf_adapter/util/npu_plugin.h
+++ b/tf_adapter/util/npu_plugin.h
@@ -50,4 +50,7 @@ int32_t GetVarAddrAndSize(const std::string &var_name, uint64_t &base_addr, uint
 
 int32_t MallocSharedMem(const ge::TensorInfo &tensor_info, uint64_t &dev_addr, uint64_t &memory_size);
 
+int32_t SetDeviceSatMode(uint32_t mode);
+
+int32_t GetDeviceSatMode();
 #endif  // TENSORFLOW_NPU_PLUGIN_H_
diff --git a/tf_adapter_2.x/npu_device/core/npu_wrapper.cpp b/tf_adapter_2.x/npu_device/core/npu_wrapper.cpp
index 9d50a7687d4529b076e0e44b4156823cb48f4cf9..4eee72f637d74c099f4a1a5c42f11ecd8413bdb6 100644
--- a/tf_adapter_2.x/npu_device/core/npu_wrapper.cpp
+++ b/tf_adapter_2.x/npu_device/core/npu_wrapper.cpp
@@ -40,6 +40,7 @@
 #include "framework/omg/parser/parser_api.h"
 #include "ge/ge_api.h"
 
+#include "acl/acl_rt.h"
 #include "npu_aoe.h"
 #include "npu_device_register.h"
 #include "npu_global.h"
@@ -119,6 +120,7 @@ const std::map<std::string, std::string> kConfigurableOptions = {
 
 namespace {
 std::unordered_set<std::string> npu_specify_ops_cache;
+constexpr uint32_t kDeviceSatModeLimit = 2U;
 }
 namespace npu {
 void ParseGlobalOptions(int device_index, const std::map<std::string, std::string> &user_options,
@@ -284,6 +286,25 @@ PYBIND11_MODULE(_npu_device_backends, m) {
               << ", it will take effect in the next training loop";
   });
 
+  (void)m.def("SetDeviceSatMode", [](uint32_t mode) {
+    if (mode > kDeviceSatModeLimit) {
+      LOG(ERROR) << "overflow mode is unvalid" << mode;
+      return;
+    }
+    aclrtSetDeviceSatMode(aclrtFloatOverflowMode(mode));
+  });
+
+  (void)m.def("GetDeviceSatMode", []() -> std::int32_t {
+    aclrtFloatOverflowMode mode = ACL_RT_OVERFLOW_MODE_UNDEF;
+    aclError ret = aclrtGetDeviceSatMode(&mode);
+    if (ret != ACL_SUCCESS) {
+      LOG(ERROR) << "get device sat mode failed";
+      return -1;
+    }
+    LOG(INFO) << "get deviceSatMode success";
+    return static_cast<int32_t>(mode);
+  });
+
   (void)m.def("RunContextOptionsSetMemoryOptimizeOptions", &RunContextOptionsSetMemoryOptimizeOptions);
   (void)m.def("CleanRunContextOptions", &CleanRunContextOptions);
   (void)m.def("RunContextOptionsSetGraphParallelOptions", &RunContextOptionsSetGraphParallelOptions);
diff --git a/tf_adapter_2.x/python/npu_device/__init__.py b/tf_adapter_2.x/python/npu_device/__init__.py
index 99207552ba1f1a416909977ec93cd6f9e595c594..1089a95ec75bd00c43c89c0dc07ff3d82d33f00c 100644
--- a/tf_adapter_2.x/python/npu_device/__init__.py
+++ b/tf_adapter_2.x/python/npu_device/__init__.py
@@ -21,6 +21,8 @@ from npu_device.npu_device import gen_npu_ops
 from npu_device.npu_device import global_options
 from npu_device.npu_device import set_npu_loop_size
 from npu_device.npu_device import npu_run_context
+from npu_device.npu_device import set_device_sat_mode
+from npu_device.npu_device import get_device_sat_mode
 
 from npu_device.utils.scope import keep_dtype_scope
 from npu_device.utils.scope import npu_recompute_scope
diff --git a/tf_adapter_2.x/python/npu_device/npu_device.py b/tf_adapter_2.x/python/npu_device/npu_device.py
index d3412cd74016759a7279034f3452a91a31b51bc7..472a810204e84ca0aae7473007bc514bf159d988 100644
--- a/tf_adapter_2.x/python/npu_device/npu_device.py
+++ b/tf_adapter_2.x/python/npu_device/npu_device.py
@@ -64,6 +64,14 @@ def set_npu_loop_size(loop_size):
     _npu_device_backends.SetNpuLoopSize(loop_size)
 
 
+def set_device_sat_mode(mode):
+    _npu_device_backends.SetDeviceSatMode(mode)
+
+
+def get_device_sat_mode():
+    return _npu_device_backends.GetDeviceSatMode()
+
+
 _global_options = None
 _global_options_lock = threading.Lock()
 
diff --git a/tf_adapter_2.x/tests/st/adapter2_st.py b/tf_adapter_2.x/tests/st/adapter2_st.py
index 55ea54778a7f04b808f69f06450575d0b4b474a4..f107b51ec611d461bd47f7b893b9f36de848526d 100644
--- a/tf_adapter_2.x/tests/st/adapter2_st.py
+++ b/tf_adapter_2.x/tests/st/adapter2_st.py
@@ -23,7 +23,8 @@ os.environ['ASCEND_OPP_PATH'] = 'non-existed-path'
 
 import npu_device
 from npu_device.npu_device import stupid_repeat
-
+from npu_device.npu_device import set_device_sat_mode
+from npu_device.npu_device import get_device_sat_mode
 import unittest
 import tensorflow as tf
 from tensorflow.python.eager import context
@@ -70,6 +71,12 @@ def foo_cpu_add_(v):
 
 
 class Adapter2St(unittest.TestCase):
+    def test_set_device_sat_mode(self):
+        set_device_sat_mode(2)
+        self.assertTrue(get_device_sat_mode(), -1);
+        set_device_sat_mode(1)
+        self.assertTrue(get_device_sat_mode(), 1);
+
     def test_mix_resource(self):
         with context.device("/job:localhost/replica:0/task:0/device:CPU:0"):
             x = tf.Variable(1)
diff --git a/tf_adapter_2.x/tests/stub/acl_stub.cpp b/tf_adapter_2.x/tests/stub/acl_stub.cpp
index 9f68453153ac8f4b2fc11bf013f981666c7b814d..8d1c7393b2d3553f00d3476fc71487d6deab7216 100644
--- a/tf_adapter_2.x/tests/stub/acl_stub.cpp
+++ b/tf_adapter_2.x/tests/stub/acl_stub.cpp
@@ -23,6 +23,11 @@ limitations under the License.
 #include "acl/acl_rt.h"
 #include "acl/acl_tdt.h"
 
+namespace {
+const uint32_t kDeviceSatModeLimit = 2U;
+std::uint32_t deviceSatMode = 2U;
+}
+
 struct aclopAttr {};
 struct aclDataBuffer {};
 struct aclTensorDesc {};
@@ -197,6 +202,22 @@ aclError acltdtReceiveTensor(const acltdtChannelHandle *handle, acltdtDataset *d
   return ACL_ERROR_NONE;
 }
 
+aclError aclrtSetDeviceSatMode(aclrtFloatOverflowMode mode) {
+  if (mode != ACL_RT_OVERFLOW_MODE_SATURATION && mode != ACL_RT_OVERFLOW_MODE_INFNAN) {
+    deviceSatMode = 2U;
+    return ACL_ERROR_INVALID_PARAM;
+  }
+  deviceSatMode = mode;
+  return ACL_ERROR_NONE;
+}
+
+aclError aclrtGetDeviceSatMode(aclrtFloatOverflowMode *mode) {
+  if (deviceSatMode >= kDeviceSatModeLimit) {
+    return ACL_ERROR_FAILURE;
+  }
+  *mode = aclrtFloatOverflowMode(deviceSatMode);
+  return ACL_ERROR_NONE;
+}
 #ifdef __cplusplus
 }
-#endif
\ No newline at end of file
+#endif