diff --git a/debug/accuracy_tools/msprobe/mindspore/monitor/optimizer_hook.py b/debug/accuracy_tools/msprobe/mindspore/monitor/optimizer_hook.py
new file mode 100644
index 0000000000000000000000000000000000000000..891d8d2666a7e56892ac1e50121e0720828341e2
--- /dev/null
+++ b/debug/accuracy_tools/msprobe/mindspore/monitor/optimizer_hook.py
@@ -0,0 +1,50 @@
+import mindspore as ms
+from mindspore import Tensor, ops
+
+ORIGIN_GRAD = ops.composite.base._Grad
+
+class CustomGradOperation:
+    def custom_grad_fn(self, gradients):
+        for idx, grad in enumerate(gradients):
+            ops.TensorDump()(rf"dump/grad", grad)
+        return gradients
+
+    def __init__(self, *args, **kwargs):
+        """Initialize CustomGradOperation"""
+        super().__init__()
+        self.grad = ops.GradOperation(*args, **kwargs)
+
+    def __call__(self, *args, **kwargs):
+        grad_fn = self.grad(*args, **kwargs)
+
+        def wrapped_grad_fn(*args, **kwargs):
+            gradients = grad_fn(*args, **kwargs)
+            self.custom_grad_fn(gradients)
+            return gradients
+        return wrapped_grad_fn
+
+
+class CustomLiteGradOperation:
+    def custom_grad_fn(self, gradients):
+        for idx, grad in enumerate(gradients):
+            ops.TensorDump()(rf"dump/grad", grad)
+        return gradients
+
+    def __init__(self, *args, **kwargs):
+        """Initialize CustomGradOperation"""
+        super().__init__()
+        self.grad = ORIGIN_GRAD(*args, **kwargs)
+
+    def __call__(self, *args, **kwargs):
+        grad_fn = self.grad(*args, **kwargs)
+
+        def wrapped_grad_fn(*args, **kwargs):
+            gradients = grad_fn(*args, **kwargs)
+            self.custom_grad_fn(gradients)
+            return gradients
+        return wrapped_grad_fn
+
+
+def enable_hook():
+    ops.composite.GradOperation = CustomGradOperation
+    ops.composite.base._Grad = CustomLiteGradOperation
\ No newline at end of file