diff --git a/tf_adapter/util/npu_attrs.cc b/tf_adapter/util/npu_attrs.cc
index e5026aaa2a5d7f8aea6831e367365e4944adf382..66de3f6bb1d5e099e59c90de876ff6045f7f43e4 100644
--- a/tf_adapter/util/npu_attrs.cc
+++ b/tf_adapter/util/npu_attrs.cc
@@ -531,7 +531,7 @@ std::map<std::string, std::string> NpuAttrs::GetSessOptions(const OpKernelConstr
 
 std::map<std::string, std::string> NpuAttrs::GetDefaultInitOptions() {
   std::map<std::string, std::string> init_options;
-  init_options["ge.exec.precision_mode"] = "allow_fp32_to_fp16";
+  init_options["ge.exec.precision_mode"] = "";
   init_options[ge::OPTION_EXEC_PROFILING_MODE] = "0";
   init_options[ge::OPTION_EXEC_PROFILING_OPTIONS] = "";
   init_options[ge::OPTION_GRAPH_RUN_MODE] = "1";
@@ -546,7 +546,7 @@ std::map<std::string, std::string> NpuAttrs::GetDefaultInitOptions() {
 }
 
 std::map<std::string, std::string> NpuAttrs::GetInitOptions(const OpKernelConstruction *ctx) {
-  std::string precision_mode = "allow_fp32_to_fp16";
+  std::string precision_mode = "";
   std::string profiling_mode = "0";
   std::string static_memory_policy = "0";
   std::string auto_tune_mode;
@@ -627,7 +627,7 @@ std::map<std::string, std::string> NpuAttrs::GetInitOptions(const OpKernelConstr
   }
 
   if (precision_mode.empty()) {
-    init_options_[ge::PRECISION_MODE] = "allow_fp32_to_fp16";
+    init_options_[ge::PRECISION_MODE] = "";
   } else {
     init_options_[ge::PRECISION_MODE] = precision_mode;
   }
@@ -1803,9 +1803,9 @@ Status NpuAttrs::SetNpuOptimizerAttr(const GraphOptimizationPassOptions &options
         NPU_REQUIRES_OK(CheckValueAllowed<std::string>(precision_mode, kPrecisionModeList));
       } else {
         if (static_cast<bool>(graph_run_mode)) {
-          precision_mode = "allow_fp32_to_fp16";
+          precision_mode = "";
         } else {
-          precision_mode = "force_fp16";
+          precision_mode = "";
         }
       }
       if (params.count("soc_config") > 0) {
@@ -2137,7 +2137,7 @@ Status NpuAttrs::SetNpuOptimizerAttr(const GraphOptimizationPassOptions &options
 
   init_options_["precision_mode"] = precision_mode;
   if (precision_mode.empty()) {
-    init_options_[ge::PRECISION_MODE] = "allow_fp32_to_fp16";
+    init_options_[ge::PRECISION_MODE] = "";
   } else {
     init_options_[ge::PRECISION_MODE] = precision_mode;
   }
diff --git a/tf_adapter_2.x/python/npu_device/configs/npu_config.py b/tf_adapter_2.x/python/npu_device/configs/npu_config.py
index 0d24e980a1c277ab9692917ed9db1d1502bd5d30..8c8cf17f001832cbc8a70c7720c15e2a4b51b31e 100644
--- a/tf_adapter_2.x/python/npu_device/configs/npu_config.py
+++ b/tf_adapter_2.x/python/npu_device/configs/npu_config.py
@@ -35,7 +35,7 @@ class NpuConfig(NpuBaseConfig):
         self.variable_format_optimize = DeprecatedValue([True, False], replacement=None)
         self.enable_scope_fusion_passes = OptionValue(None, None)
         self.fusion_switch_file = OptionValue(None, None)
-        self.precision_mode = OptionValue('allow_fp32_to_fp16',
+        self.precision_mode = OptionValue('',
                                           ['force_fp32', 'allow_fp32_to_fp16', 'force_fp16', 'must_keep_origin_dtype',
                                            'allow_mix_precision', 'cube_fp16in_fp32out', 'allow_mix_precision_fp16',
                                            'allow_mix_precision_bf16', 'allow_fp32_to_bf16'])