revert: enable return_outputs when necessary

Author: cwher

examples/images/vit/vit_train_demo.py

@@ -37,7 +37,7 @@ 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=False
+            data_iter, model, criterion, optimizer, return_loss=True, return_outputs=True
         )
         loss, outputs = output_dict["loss"], output_dict["outputs"]
     else:

examples/language/bert/finetune.py

@@ -70,7 +70,7 @@ def evaluate_subset(dataloader: DataLoader):
                 current_rank = dist.get_rank()
                 batch = iter([batch])
 
-                outputs = booster.execute_pipeline(batch, model, criterion, return_loss=True, return_outputs=False)
+                outputs = booster.execute_pipeline(batch, model, criterion, return_loss=True, return_outputs=True)
 
                 if is_pp_last_device:
                     logits = outputs["outputs"]["logits"]

examples/language/gpt/hybridparallelism/finetune.py

@@ -64,7 +64,7 @@ def evaluate_subset(dataloader: DataLoader):
                 current_pp_group_ranks = pg_mesh.get_ranks_in_group(pp_group)
                 current_rank = dist.get_rank()
                 batch = iter([batch])
-                outputs = booster.execute_pipeline(batch, model, criterion, return_loss=True, return_outputs=False)
+                outputs = booster.execute_pipeline(batch, model, criterion, return_loss=True, return_outputs=True)
 
                 if is_pp_last_stage:
                     logits = outputs["outputs"]["logits"]

tests/test_shardformer/test_model/_utils.py

@@ -173,7 +173,7 @@ def _criterion(outputs, inputs):
 
         data_iter = iter([data])
         sharded_output = booster.execute_pipeline(
-            data_iter, sharded_model, _criterion, sharded_optimizer, return_loss=True, return_outputs=False
+            data_iter, sharded_model, _criterion, sharded_optimizer, return_loss=True, return_outputs=True
         )
         sharded_loss = sharded_output["loss"]
     else: