diff --git a/torch_npu/csrc/core/npu/NPUGraph.cpp b/torch_npu/csrc/core/npu/NPUGraph.cpp
index a00448bd1cf547cb2f641eb47ff940c902341610..503a6c591cc2d8c1b660ceb903b50d08add9635f 100644
--- a/torch_npu/csrc/core/npu/NPUGraph.cpp
+++ b/torch_npu/csrc/core/npu/NPUGraph.cpp
@@ -97,7 +97,7 @@ NPUGraph::NPUGraph()
 
 void NPUGraph::capture_begin(MempoolId_t pool, aclmdlRICaptureMode capture_mode)
 {
-    static const auto _task_queue_enable = c10_npu::option::OptionsManager::GetTaskQueueEnable();
+    const auto _task_queue_enable = c10_npu::option::OptionsManager::GetTaskQueueEnable();
     TORCH_CHECK(_task_queue_enable != 2,
         "Do not support TASK_QUEUE_ENABLE = 2 during NPU graph capture, please "
         "export TASK_QUEUE_ENABLE=1/0.",
diff --git a/torch_npu/csrc/core/npu/NPUQueue.cpp b/torch_npu/csrc/core/npu/NPUQueue.cpp
index 579514ab37390f36aa208e7711c6fcec131a9f98..98b807a9a47141437ad64ce53e1aa1e1397190e0 100644
--- a/torch_npu/csrc/core/npu/NPUQueue.cpp
+++ b/torch_npu/csrc/core/npu/NPUQueue.cpp
@@ -388,7 +388,7 @@ void Repository::CheckDeviceError(int ret, std::string& err_msg)
 bool Repository::ReadQueue()
 {
     if (IsEmptyQueue()) {
-        static const auto task_queue_enable = c10_npu::option::OptionsManager::GetTaskQueueEnable();
+        const auto task_queue_enable = c10_npu::option::OptionsManager::GetTaskQueueEnable();
         if (task_queue_enable == 2) {
             // read queue polls for at most 1 ms when queue is empty.
             for (int i = 0; i < READ_QUEUE_POLL_MAX_LOOP; ++i) {
diff --git a/torch_npu/csrc/core/npu/register/OptionsManager.cpp b/torch_npu/csrc/core/npu/register/OptionsManager.cpp
index 1eedf54559d1171ab270b1288523fdf0295ffa63..ba79866cc063c7efac4ede2e3c5d3798cadd2ce6 100644
--- a/torch_npu/csrc/core/npu/register/OptionsManager.cpp
+++ b/torch_npu/csrc/core/npu/register/OptionsManager.cpp
@@ -12,6 +12,7 @@
 #include "torch_npu/csrc/core/npu/register/OptionRegister.h"
 #include "torch_npu/csrc/core/npu/register/OptionsManager.h"
 #include "torch_npu/csrc/core/npu/NPUFunctions.h"
+#include "torch_npu/csrc/core/npu/NPUStream.h"
 #include "torch_npu/csrc/core/npu/NPUCachingAllocator.h"
 #include "torch_npu/csrc/npu/memory_snapshot.h"
 
@@ -512,23 +513,35 @@ char* OptionsManager::GetCpuAffinityConf()
     return std::getenv("CPU_AFFINITY_CONF");
 }
 
+static uint32_t task_queue_enable = []() -> uint32_t {
+    char* env_val = std::getenv("TASK_QUEUE_ENABLE");
+    int64_t task_queue_enable = (env_val != nullptr) ? strtol(env_val, nullptr, 10) : 1;
+    std::unordered_map<int32_t, std::string> taskQueueEnableMode = getTaskQueueEnableMode();
+    if (taskQueueEnableMode.find(task_queue_enable) == taskQueueEnableMode.end()) {
+        TORCH_CHECK(false, "TASK_QUEUE_ENABLE should be 0, 1 or 2", PTA_ERROR(ErrCode::VALUE));
+    }
+    return static_cast<uint32_t>(task_queue_enable);
+}();
+
 uint32_t OptionsManager::GetTaskQueueEnable()
 {
     if (CheckBlockingEnable()) {
         return 0;
     }
-    const static uint32_t task_queue_enable = []() -> uint32_t {
-        char* env_val = std::getenv("TASK_QUEUE_ENABLE");
-        int64_t task_queue_enable = (env_val != nullptr) ? strtol(env_val, nullptr, 10) : 1;
-        std::unordered_map<int32_t, std::string> taskQueueEnableMode = getTaskQueueEnableMode();
-        if (taskQueueEnableMode.find(task_queue_enable) == taskQueueEnableMode.end()) {
-            TORCH_CHECK(false, "TASK_QUEUE_ENABLE should be 0, 1 or 2", PTA_ERROR(ErrCode::VALUE));
-        }
-        return static_cast<uint32_t>(task_queue_enable);
-    }();
     return task_queue_enable;
 }
 
+void OptionsManager::SetTaskQueueEnable(uint32_t value)
+{
+    if (task_queue_enable != value) {
+        NPUStatus ret = c10_npu::emptyAllNPUStream();
+        if (ret != SUCCESS) {
+            ASCEND_LOGE("Failed to empty NPU task queue, ret: %s", ret.c_str());
+        }
+        task_queue_enable = value;
+    }
+}
+
 bool OptionsManager::CheckForceUncached()
 {
     const static bool force_uncached = []() -> bool {
diff --git a/torch_npu/csrc/core/npu/register/OptionsManager.h b/torch_npu/csrc/core/npu/register/OptionsManager.h
index 73f5dbcb81f9fc268d8ef9122407e66b976dad08..f694f46ca4b651df5d0d5196624d0a075153f865 100644
--- a/torch_npu/csrc/core/npu/register/OptionsManager.h
+++ b/torch_npu/csrc/core/npu/register/OptionsManager.h
@@ -126,6 +126,7 @@ public:
     static uint32_t GetHcclBufferSize();
     static uint32_t GetP2PBufferSize();
     static uint32_t GetTaskQueueEnable();
+    static void SetTaskQueueEnable(uint32_t value);
     static uint32_t GetAclOpInitMode();
     static uint32_t GetStreamsPerDevice();
     static char* GetCpuAffinityConf();
diff --git a/torch_npu/csrc/npu/Module.cpp b/torch_npu/csrc/npu/Module.cpp
index 24ed076429f20328e132af42c87346042cecf87a..6bbe1e6537738911b1e59a63aae69ae2e1cbda4b 100644
--- a/torch_npu/csrc/npu/Module.cpp
+++ b/torch_npu/csrc/npu/Module.cpp
@@ -29,6 +29,7 @@
 #include "torch_npu/csrc/core/npu/NPUAffinityController.h"
 #include "torch_npu/csrc/core/npu/NPUGuard.h"
 #include "torch_npu/csrc/core/npu/NpuVariables.h"
+#include "torch_npu/csrc/core/npu/register/OptionsManager.h"
 #include "torch_npu/csrc/core/npu/sys_ctrl/npu_sys_ctrl.h"
 #include "torch_npu/csrc/core/npu/register/OptionRegister.h"
 #include "torch_npu/csrc/core/OverflowUtils.h"
@@ -1661,6 +1662,18 @@ static PyObject* THNPModule_is_gte_cann_version(PyObject* self, PyObject *args)
     END_HANDLE_TH_ERRORS
 }
 
+PyObject* THNPModule_set_task_queue_enable(PyObject* self, PyObject* args)
+{
+    HANDLE_TH_ERRORS
+    int value;
+    if (!PyArg_ParseTuple(args, "i", &value)) {
+        throw torch::TypeError("Pybind failed to parse parameters." + PTA_ERROR(ErrCode::TYPE));
+    }
+    c10_npu::option::OptionsManager::SetTaskQueueEnable(static_cast<uint32_t>(value));
+    Py_RETURN_NONE;
+    END_HANDLE_TH_ERRORS
+}
+
 
 static struct PyMethodDef THNPModule_methods[] = {
     {"_npu_init", (PyCFunction)THNPModule_initExtension, METH_NOARGS, nullptr},
@@ -1723,6 +1736,7 @@ static struct PyMethodDef THNPModule_methods[] = {
     {"_npu_clear_fft_plan_cache", (PyCFunction)THNPModule_npu_clear_fft_plan_cache, METH_NOARGS, nullptr},
     {"_get_cann_version", (PyCFunction)THNPModule_get_cann_version, METH_O, nullptr},
     {"_is_gte_cann_version", (PyCFunction)THNPModule_is_gte_cann_version, METH_VARARGS, nullptr},
+    {"_set_task_queue_enable", (PyCFunction)THNPModule_set_task_queue_enable, METH_VARARGS, nullptr},
     {nullptr}};
 
 TORCH_NPU_API PyMethodDef* THNPModule_get_methods()