[hotfix] set return_outputs=False in examples and polish code (#5404)

* fix: simplify merge_batch

* fix: use return_outputs=False to eliminate extra memory consumption

* feat: add return_outputs warning

* style: remove `return_outputs=False` as it is the default value

docs/source/en/advanced_tutorials/train_gpt_using_hybrid_parallelism.md

@@ -178,7 +178,7 @@ def train_epoch(
         for _ in pbar:
             if use_pipeline:
                 outputs = booster.execute_pipeline(
-                    train_dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True
+                    train_dataloader_iter, model, _criterion, optimizer, return_loss=True
                 )
                 # Backward and optimize
                 if is_pp_last_stage:

docs/source/en/advanced_tutorials/train_vit_with_hybrid_parallelism.md

@@ -231,7 +231,7 @@ def run_forward_backward(
     if isinstance(booster.plugin, HybridParallelPlugin) and booster.plugin.pp_size > 1:
         # run pipeline forward backward when enabling pp in hybrid parallel plugin
         output_dict = booster.execute_pipeline(
-            data_iter, model, criterion, optimizer, return_loss=True, return_outputs=True
+            data_iter, model, criterion, optimizer, return_loss=True
         )
         loss, outputs = output_dict["loss"], output_dict["outputs"]
     else:

docs/source/en/features/pipeline_parallel.md

@@ -198,8 +198,7 @@ def train_epoch(epoch: int, model: nn.Module, optimizer: Optimizer, _criterion:
                                                 model,
                                                 _criterion,
                                                 optimizer,
-                                                return_loss=True,
-                                                return_outputs=True)
+                                                return_loss=True)
             # Backward and optimize
             if is_pp_last_stage:
                 loss = outputs['loss']

docs/source/en/features/shardformer.md

@@ -271,7 +271,7 @@ However, if pipeline parallel is enabled, there are several usages different fro
 3. Do forward and backward passing through calling `Booster.execute_pipeline` method:
     ```python
     outputs = booster.execute_pipeline(
-        train_dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True
+        train_dataloader_iter, model, _criterion, optimizer, return_loss=True
     )
     ```
     Backward passing has been completed by this method, so there is no need to call `loss.backward()` after executing this method.

docs/source/zh-Hans/advanced_tutorials/train_gpt_using_hybrid_parallelism.md

@@ -175,7 +175,7 @@ def train_epoch(
         for _ in pbar:
             if use_pipeline:
                 outputs = booster.execute_pipeline(
-                    train_dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True
+                    train_dataloader_iter, model, _criterion, optimizer, return_loss=True
                 )
                 # Backward and optimize
                 if is_pp_last_stage:

docs/source/zh-Hans/advanced_tutorials/train_vit_with_hybrid_parallelism.md

@@ -234,7 +234,7 @@ def run_forward_backward(
     if isinstance(booster.plugin, HybridParallelPlugin) and booster.plugin.pp_size > 1:
         # run pipeline forward backward when enabling pp in hybrid parallel plugin
         output_dict = booster.execute_pipeline(
-            data_iter, model, criterion, optimizer, return_loss=True, return_outputs=True
+            data_iter, model, criterion, optimizer, return_loss=True
         )
         loss, outputs = output_dict["loss"], output_dict["outputs"]
     else:

docs/source/zh-Hans/features/pipeline_parallel.md

@@ -193,8 +193,7 @@ def train_epoch(epoch: int, model: nn.Module, optimizer: Optimizer, _criterion:
                                                 model,
                                                 _criterion,
                                                 optimizer,
-                                                return_loss=True,
-                                                return_outputs=True)
+                                                return_loss=True)
             # Backward and optimize
             if is_pp_last_stage:
                 loss = outputs['loss']

docs/source/zh-Hans/features/shardformer.md

@@ -264,7 +264,7 @@ elif args.plugin == "hybrid_parallel":
 3. 通过调用`Booster.execute_pipeline` 方法来执行前向和后向传递:
     ```python
     outputs = booster.execute_pipeline(
-        train_dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True
+        train_dataloader_iter, model, _criterion, optimizer, return_loss=True
     )
     ```
     该方法会自动执行后向传递，所以在执行该方法后不需要再调用 `loss.backward()`方法。