[Dy2static] [bugfix] fixed a bug which happens while parsing grad var name (#52110)

* fix dy2s grad name parse

* pre-commit

* bug fix

* Fix grad/ error

* Format code

---------
Co-authored-by: N0x45f <wangzhen45@baidu.com>

python/paddle/fluid/tests/unittests/dygraph_to_static/test_gradname_parse.py

+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import unittest
+
+import paddle
+from paddle import ParamAttr
+from paddle.nn import BatchNorm, Linear
+
+
+class SimpleNet(paddle.nn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.linear0 = Linear(100, 50)
+        self.linear1 = Linear(50, 10)
+
+        param_attr0 = ParamAttr(name="aaaprefix_bn_scale")
+        bias_attr0 = ParamAttr(name="aaaprefix_bn_offset")
+        self.bn0 = BatchNorm(50, param_attr=param_attr0, bias_attr=bias_attr0)
+
+        param_attr1 = ParamAttr(name="bn_scale")
+        bias_attr1 = ParamAttr(name="bn_offset")
+        self.bn1 = BatchNorm(10, param_attr=param_attr1, bias_attr=bias_attr1)
+
+    def forward(self, x):
+        x1 = self.linear0(x)
+        x2 = self.bn0(x1)
+        x3 = self.linear1(x2)
+        x4 = self.bn1(x3)
+        dx = paddle.grad(x4, x)
+        return dx[0]
+
+
+class TestGradNameParse(unittest.TestCase):
+    def test_grad_name_parse(self):
+        net = SimpleNet()
+        opt = paddle.optimizer.Adam(
+            learning_rate=0.1,
+            parameters=net.parameters(),
+            weight_decay=paddle.regularizer.L1Decay(0.01),
+        )
+        net = paddle.jit.to_static(net)
+        inp = paddle.rand([100, 100], dtype="float32")
+        inp.stop_gradient = False
+        out = net(inp)
+        loss = out.mean()
+        loss.backward()
+
+        for name, param in net.bn1.named_parameters():
+            if name in ["bn_scale", "bn_offset"]:
+                assert param.shape == param.grad.shape
+
+        opt.minimize(loss)
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/jit/dy2static/utils.py

@@ -1454,16 +1454,19 @@ def _param_grad_names(program_desc, params):
     # NOTE: `names` and `params` must be in the same order so that
     # the param grad name can be set correctly in the run_program.
     for param in params:
-        candidate = [
-            var.name()
-            for var in program_desc.block(0).all_vars()
-            if var.name().endswith(param.name + '@GRAD')
-        ]
+        candidate = []
+        suffix = param.name + '@GRAD'
+        for var in program_desc.block(0).all_vars():
+            var_name = var.name()
+            if var_name.endswith(suffix):
+                prefix_count = var_name.count('grad/')
+                if 'grad/' * prefix_count + suffix == var_name:
+                    candidate.append(var_name)
+
         if candidate:
             names.append(max(candidate, key=lambda name: name.count('grad/')))
         else:
-            names.append(param.name + '@GRAD')
-
+            names.append(suffix)
     return names