diff --git a/debug/accuracy_tools/grad_tool/common/utils.py b/debug/accuracy_tools/grad_tool/common/utils.py
index cdce3fda7e38940c293849b63f139ce96c5d1d55..2927f2b5d1fb600d2e5af141a843c35fefc0d8ea 100644
--- a/debug/accuracy_tools/grad_tool/common/utils.py
+++ b/debug/accuracy_tools/grad_tool/common/utils.py
@@ -2,6 +2,7 @@ import os
 import re
 import sys
 import time
+import shutil
 import yaml
 
 import pandas as pd
@@ -220,3 +221,15 @@ def change_mode(path, mode):
     except PermissionError as ex:
         print_error_log(f'Failed to change {path} authority. {str(ex)}')
         raise ex
+
+def remove_path(path):
+    if not os.path.exists(path):
+        return
+    try:
+        if os.path.islink(path) or os.path.isfile(path):
+            os.remove(path)
+        else:
+            shutil.rmtree(path)
+    except PermissionError as err:
+        print_error_log("Failed to delete {}. Please check the permission.".format(path))
+        raise err
\ No newline at end of file
diff --git a/debug/accuracy_tools/grad_tool/grad_pt/grad_monitor.py b/debug/accuracy_tools/grad_tool/grad_pt/grad_monitor.py
index f3079e622c29eefe20a6e1fdc9d372002b596610..eb7b3a896dfe5269c3b14673389285b4474cfad9 100644
--- a/debug/accuracy_tools/grad_tool/grad_pt/grad_monitor.py
+++ b/debug/accuracy_tools/grad_tool/grad_pt/grad_monitor.py
@@ -6,7 +6,7 @@ from torch.optim.optimizer import register_optimizer_step_pre_hook
 from grad_tool.common.base_monitor import BaseMonitor
 from grad_tool.grad_pt.grad_stat_csv import GradStatCsv
 from grad_tool.common.utils import check_numeral_list_ascend, data_in_list_target, \
-    write_csv, print_info_log, create_directory, print_warn_log, change_mode
+    write_csv, print_info_log, create_directory, print_warn_log, change_mode, remove_path
 from grad_tool.grad_pt.utils import get_rank_id, print_rank_0, GradConst
 
 
@@ -49,6 +49,8 @@ class PtGradientMonitor(BaseMonitor):
             print_warn_log(f"the file in {self._output_path} will be recoverd")
         self._step = -1
         self._param2name = defaultdict(str)
+        self._exec_order = []
+        self._recording_exec_order = True
 
     @property
     def output_path(self):
@@ -65,10 +67,18 @@ class PtGradientMonitor(BaseMonitor):
         change_mode(save_filepath, 0o640)
 
     def monitor(self, model):
+
+        def tensor_hook(name):
+            def record_order(grad):
+                if self._recording_exec_order:
+                    self._exec_order.append(name)
+            return record_order
+
         print_rank_0("> parameter names:")
         for name, param in model.named_parameters():
             self._param2name[param] = name
             print_rank_0(f"\t{name}")
+            param.register_hook(tensor_hook(name))
         setattr(self, "_rank", get_rank_id())
         if torch.distributed.is_initialized() and not data_in_list_target(getattr(self, "_rank"), self._target_ranks):
             return
@@ -79,8 +89,13 @@ class PtGradientMonitor(BaseMonitor):
             raise AttributeError("grad monitor need attribute {_rank}")
         return not torch.distributed.is_initialized() or data_in_list_target(getattr(self, "_rank"), self._target_ranks)
 
+    def _output_sort_by_exec_order(self, output_lines):
+        dict_by_name = {x[0]: x for x in output_lines}
+        return [dict_by_name[i] for i in self._exec_order]
+
     def _hook_optimizer(self):
         def optimizer_pre_step_hook(optimizer, args, kargs):
+            self._recording_exec_order = False
             self._step += 1
             if not data_in_list_target(self._step, self._target_step):
                 return
@@ -97,8 +112,16 @@ class PtGradientMonitor(BaseMonitor):
                 if self._level_adp["have_grad_direction"]:
                     PtGradientMonitor.save_grad_direction(param_name, grad,
                                                     f'{self._output_path}/rank{self._rank}/step{self._step}')
+                    
+            if self._exec_order:
+                if len(self._exec_order) != len(output_lines):
+                    print_warn_log(f"grad_number in tensor_hook not equal to grad_number in optimizer_hook: \
+                                   {len(self._exec_order)} vs {len(output_lines)}. maybe something wrong happend.")
+                else:
+                    output_lines = self._output_sort_by_exec_order(output_lines)
             output_path = os.path.join(self._output_path, f"rank{getattr(self, '_rank')}",
                                        f"grad_summary_{self._step}.csv")
+            remove_path(output_path)
             write_csv(output_path, output_lines,
                       GradStatCsv.generate_csv_header(self._level_adp, self._bounds))
 
diff --git a/debug/accuracy_tools/grad_tool/test/ut/test_grad_monitor.py b/debug/accuracy_tools/grad_tool/test/ut/test_grad_monitor.py
index 77a51e0941fa80edd53c2814c9c1a9252001b9ae..2121db7cc586074984f1dd6af17a7379c440e3ec 100644
--- a/debug/accuracy_tools/grad_tool/test/ut/test_grad_monitor.py
+++ b/debug/accuracy_tools/grad_tool/test/ut/test_grad_monitor.py
@@ -64,4 +64,4 @@ class TestGradMonitor(unittest.TestCase):
         self.assertEqual(len(items), 1)
         with open(os.path.join(compare_output_path, items[0], "similarities.csv"), 'r') as f:
             data = f.read()
-        self.assertEqual(hashlib.md5(data.encode("utf-8")).hexdigest(), "0e2bd636b48245d387647523c8517982")
+        self.assertEqual(hashlib.md5(data.encode("utf-8")).hexdigest(), "a968d0cc1c6e289c00ff4fd5811811ca")