diff --git a/deepmd/pd/model/descriptor/dpa3.py b/deepmd/pd/model/descriptor/dpa3.py
index 95fe69422b..de6bc5d45f 100644
--- a/deepmd/pd/model/descriptor/dpa3.py
+++ b/deepmd/pd/model/descriptor/dpa3.py
@@ -255,11 +255,14 @@ def init_subclass_params(sub_data: dict | Any, sub_class: type) -> Any:
             "buffer_ntypes", paddle.to_tensor(self.ntypes, dtype="int64")
         )
 
-        # set trainable
-        for param in self.parameters():
-            param.requires_grad = trainable
+        self._apply_trainable()
         self.compress = False
 
+    def _apply_trainable(self) -> None:
+        """Apply the descriptor-level trainable setting to every parameter."""
+        for param in self.parameters():
+            param.stop_gradient = not self.trainable
+
     def get_rcut(self) -> float:
         """Returns the cut-off radius."""
         return self.rcut
@@ -557,6 +560,9 @@ def t_cvt(xx: Any) -> paddle.Tensor:
         obj.repflows.layers = paddle.nn.LayerList(
             [RepFlowLayer.deserialize(layer) for layer in repflow_layers]
         )
+        # Deserialization replaces several sublayers after construction, so apply
+        # the descriptor-level setting again to their newly registered parameters.
+        obj._apply_trainable()
         return obj
 
     def forward(
diff --git a/source/tests/pd/model/test_dpa3.py b/source/tests/pd/model/test_dpa3.py
index 7b3c8ba18f..fef0da2d03 100644
--- a/source/tests/pd/model/test_dpa3.py
+++ b/source/tests/pd/model/test_dpa3.py
@@ -1,6 +1,9 @@
 # SPDX-License-Identifier: LGPL-3.0-or-later
 import itertools
 import unittest
+from unittest.mock import (
+    patch,
+)
 
 import numpy as np
 import paddle
@@ -36,6 +39,49 @@ class TestDescrptDPA3(unittest.TestCase, TestCaseSingleFrameWithNlist):
     def setUp(self) -> None:
         TestCaseSingleFrameWithNlist.setUp(self)
 
+    def test_non_trainable_parameters_stop_gradient(self) -> None:
+        """A non-trainable descriptor must disable Paddle autograd parameters."""
+        # Paddle 3.0 does not expose the newer ``requires_grad`` compatibility
+        # alias. Ignore writes to that alias so this test retains minimum-version
+        # semantics when the CI environment uses a newer Paddle release.
+        requires_grad_alias = property(
+            lambda parameter: not parameter.stop_gradient,
+            lambda _parameter, _value: None,
+        )
+        with patch.object(
+            paddle.Tensor, "requires_grad", requires_grad_alias, create=True
+        ):
+            descriptor = DescrptDPA3(
+                self.nt,
+                repflow=RepFlowArgs(
+                    n_dim=4,
+                    e_dim=4,
+                    a_dim=4,
+                    nlayers=1,
+                    e_sel=2,
+                    a_sel=1,
+                    axis_neuron=2,
+                ),
+                trainable=False,
+                add_chg_spin_ebd=True,
+                seed=GLOBAL_SEED,
+            )
+            deserialized_descriptor = DescrptDPA3.deserialize(descriptor.serialize())
+
+        for stage, checked_descriptor in (
+            ("constructed", descriptor),
+            ("deserialized", deserialized_descriptor),
+        ):
+            with self.subTest(stage=stage):
+                parameters = list(checked_descriptor.named_parameters())
+                self.assertTrue(parameters)
+                parameters_with_grad = [
+                    name
+                    for name, parameter in parameters
+                    if not parameter.stop_gradient
+                ]
+                self.assertEqual([], parameters_with_grad)
+
     def test_consistency(
         self,
     ) -> None: