[AutoParallel] keep lr_sheduler same bewteen executor and engine (#55516)

python/paddle/distributed/auto_parallel/static/callbacks.py

@@ -171,14 +171,14 @@ class LRSchedulerAuto(LRScheduler):
 
         if self.by_step and self.train_step % self.acc_step == 0:
             if (
-                self.model._optimizer
-                and hasattr(self.model._optimizer, '_learning_rate')
+                self.model.optimizer
+                and hasattr(self.model.optimizer, '_learning_rate')
                 and isinstance(
-                    self.model._optimizer._learning_rate,
+                    self.model.optimizer._learning_rate,
                     paddle.optimizer.lr.LRScheduler,
                 )
             ):
-                self.model._optimizer._learning_rate.step()
+                self.model.optimizer._learning_rate.step()
 
 
 class History(Callback):

python/paddle/distributed/auto_parallel/static/engine.py

@@ -970,7 +970,9 @@ class Engine:
             save_dir=save_dir,
             verbose=verbose,
             metrics=self._metrics_name(),
-            acc_step=self._acc_steps,
+            acc_step=1
+            if self._strategy.pipeline.enable
+            else self._acc_steps,  # lr update once every local batch
         )
 
         cbks.on_begin('train')

python/paddle/distributed/auto_parallel/static/parallelizer_v2.py

@@ -223,10 +223,15 @@ class Parallelizer:
     def _generate_optimizer(
         self, main_program, startup_program, optimizer, params_grads
     ):
-        # NOTE: `apply_gradients` will add an Accumulator for a parameter only once,
+        # NOTE:
+        # 1. `apply_gradients` will add an Accumulator for a parameter only once,
         #    but optimizer will be called repeatedly in re-launch, so optimizer need to be copied.
+        # 2. lr_scheduler cannot be deepcopy, cause 'deepcopy' will lead to difference of learning_rate between executor and engine.
+        learning_rate = optimizer._learning_rate
         optimizer = copy.deepcopy(optimizer)
         self._dist_context._serial_optimizer = optimizer
+        self._dist_context._serial_optimizer._learning_rate = learning_rate
+
         with program_guard(main_program, startup_program):
             with unique_name.guard("opt_"):
                 optimizer_ops = optimizer.apply_gradients(params_grads)