diff --git a/torch_npu/csrc/core/npu/NPUAffinityController.cpp b/torch_npu/csrc/core/npu/NPUAffinityController.cpp
index 36e585f50a4c2b67f21e3d0c5ea85724453b18e8..56604d9935aed2dc924bd3a36913e79488171ea6 100644
--- a/torch_npu/csrc/core/npu/NPUAffinityController.cpp
+++ b/torch_npu/csrc/core/npu/NPUAffinityController.cpp
@@ -18,6 +18,7 @@ static thread_local ThreadType local_thread = ThreadType::MAIN_THREAD;
 static pthread_t main_thread;
 static bool start_main_thread_bind = false;
 static std::mutex core_map_mutex;
+static bool lazy_bind = true;
 
 using ThreadCoreMap = std::unordered_map<ThreadType, CoreIdRange>;
 
@@ -89,6 +90,15 @@ void parseCPUAffinityConf(uint32_t &mode, std::vector<CoreIdRange> &ranges)
         }
     }
 
+    std::regex pattern_for_lazy_bind("lazy_bind:(\\d)");
+    std::smatch match_for_lazy_bind;
+    if (std::regex_search(inputStr, match_for_lazy_bind, pattern_for_lazy_bind)) {
+        int lazy_bind_val = std::stoi(match_for_lazy_bind[1].str());
+        if (lazy_bind_val == 0) {
+            lazy_bind = false;
+        }
+    }
+
     // Handle cases where only `mode` is provided, or `mode:` without value
     if (isAllDigits(inputStr)) {
         mode = static_cast<uint32_t>(std::stoi(inputStr));
@@ -335,6 +345,14 @@ bool NeedMainThreadBind()
     return start_main_thread_bind && (local_thread == ThreadType::MAIN_THREAD);
 }
 
+bool SetThreadAffinityInInitialize()
+{
+    if (needToSetThreadAffinity() && !lazy_bind) {
+        return true;
+    }
+    return false;
+}
+
 void StartMainThreadBind(c10::DeviceIndex device_id)
 {
     if (!needToSetThreadAffinity() || local_thread == ThreadType::USER_THREAD) {
diff --git a/torch_npu/csrc/core/npu/NPUAffinityController.h b/torch_npu/csrc/core/npu/NPUAffinityController.h
index e850a47b67f3484ffafb56f0b4cc67b0eea0c0ee..5a38f885d4d68a6a6797bdc94e0a11c25c01b571 100644
--- a/torch_npu/csrc/core/npu/NPUAffinityController.h
+++ b/torch_npu/csrc/core/npu/NPUAffinityController.h
@@ -26,6 +26,7 @@ void SetThreadAffinity(int core_start, int core_end);
 
 void SetMainThread();
 bool NeedMainThreadBind();
+bool SetThreadAffinityInInitialize();
 void StartMainThreadBind(c10::DeviceIndex device_id);
 
 } // namespace c10_npu
\ No newline at end of file
diff --git a/torch_npu/csrc/core/npu/sys_ctrl/npu_sys_ctrl.cpp b/torch_npu/csrc/core/npu/sys_ctrl/npu_sys_ctrl.cpp
index 4b6707b8495b9a89c500e06f160b42950e9ae6fb..5f348e029ca05ecc337e3e68befd3e6c44967168 100644
--- a/torch_npu/csrc/core/npu/sys_ctrl/npu_sys_ctrl.cpp
+++ b/torch_npu/csrc/core/npu/sys_ctrl/npu_sys_ctrl.cpp
@@ -190,6 +190,9 @@ NpuSysCtrl::SysStatus NpuSysCtrl::Initialize(int device_id)
     lazy_fn_.clear();
 
     SetMainThread();
+    if (SetThreadAffinityInInitialize()) {
+        SetThreadAffinity(device_id_);
+    }
 
     init_flag_ = true;
     ASCEND_LOGD("Npu sys ctrl initialize successfully.");