[Dy2Stat]Set buff.persistable=False when it's not initialized (#28749)

python/paddle/fluid/dygraph/base.py

@@ -79,8 +79,12 @@ def param_guard(parameters):
                     # `mask` Tensor or `hidden_0` in RNN layers, which is equivalent to a Parameter
                     # and necessary for inferring. It will be pruned if it's not necessary for inferring.
                     else:
+                        # But if its shape is empty while created from `create_variable()`, we consider this buffer
+                        # non-persistable. See case of `drop_state` in lstm api.
+                        is_persistable = len(var_base.shape) > 0
+
                         new_var = var_base._to_static_var(
-                            to_parameter=False, persistable=True)
+                            to_parameter=False, persistable=is_persistable)
                 parameters[name] = new_var
         yield
         parameters.update(origin_parameters)

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

@@ -61,25 +61,26 @@ class TestLstm(unittest.TestCase):
             msg='dygraph_out is {}\n static_out is \n{}'.format(dygraph_out,
                                                                 static_out))
 
-    def test_save_in_eval(self):
+    def test_save_in_eval(self, with_training=True):
         paddle.jit.ProgramTranslator().enable(True)
         net = Net(12, 2)
         x = paddle.randn((2, 10, 12))
-        x.stop_gradient = False
-        dygraph_out = net(x)
-        loss = paddle.mean(dygraph_out)
-        sgd = paddle.optimizer.SGD(learning_rate=0.001,
-                                   parameters=net.parameters())
-        loss.backward()
-        sgd.step()
+        if with_training:
+            x.stop_gradient = False
+            dygraph_out = net(x)
+            loss = paddle.mean(dygraph_out)
+            sgd = paddle.optimizer.SGD(learning_rate=0.001,
+                                       parameters=net.parameters())
+            loss.backward()
+            sgd.step()
         # switch eval mode firstly
         net.eval()
         x = paddle.randn((2, 10, 12))
-        dygraph_out = net(x)
-        dropout_out = net(x)
         net = paddle.jit.to_static(
             net, input_spec=[paddle.static.InputSpec(shape=[-1, 10, 12])])
         paddle.jit.save(net, 'simple_lstm')
+
+        dygraph_out = net(x)
         # load saved model
         load_net = paddle.jit.load('simple_lstm')
 
@@ -96,6 +97,9 @@ class TestLstm(unittest.TestCase):
             msg='dygraph_out is {}\n static_out is \n{}'.format(dygraph_out,
                                                                 train_out))
 
+    def test_save_without_training(self):
+        self.test_save_in_eval(with_training=False)
+
 
 class LinearNet(nn.Layer):
     def __init__(self):