[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

applications/ColossalMoE/train.py

@@ -238,7 +238,6 @@ def main():
                         lambda x, y: x.loss,
                         optimizer,
                         return_loss=True,
-                        return_outputs=True,
                     )
                     # Backward and optimize
                     if is_pp_last_stage:

colossalai/booster/plugin/hybrid_parallel_plugin.py

@@ -1183,6 +1183,9 @@ class HybridParallelPlugin(PipelinePluginBase):
     ) -> dict:
         assert self.enable_pipeline_parallelism, "pipeline parallelism is not enabled"
 
+        if return_outputs:
+            warnings.warn("return_outputs may lead to significant extra memory consumption.")
+
         # Create a context for gradient synchronization based on the optimizer type.
         # If it's a HybridParallelZeroOptimizer, use optimizer.no_sync(); otherwise, use model.no_sync().
         # This is to avoid redundant gradient reduction in pipeline parallelism (multiple microbatch values should be reduced once),

colossalai/pipeline/schedule/one_f_one_b.py

@@ -7,7 +7,7 @@ from torch.nn import Module
 from torch.utils._pytree import tree_map
 
 from colossalai.accelerator import get_accelerator
-from colossalai.interface import ModelWrapper, OptimizerWrapper
+from colossalai.interface import OptimizerWrapper
 from colossalai.pipeline.p2p import PipelineP2PCommunication, create_send_metadata
 from colossalai.pipeline.stage_manager import PipelineStageManager
 from colossalai.utils import get_current_device
@@ -327,9 +327,7 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
             self.send_forward(output_obj)
 
         if outputs is not None:
-            if isinstance(model, ModelWrapper):
-                model = model.unwrap()
-            outputs = merge_batch(outputs, getattr(model, "batch_size_dim", 0))
+            outputs = merge_batch(outputs)
         return {"loss": accum_loss, "outputs": outputs}
 
     def run_forward_backward(
@@ -412,9 +410,7 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
         assert all(len(v) == 0 for v in input_objs) and all(len(v) == 0 for v in output_objs)
 
         if outputs is not None:
-            if isinstance(model, ModelWrapper):
-                model = model.unwrap()
-            outputs = merge_batch(outputs, getattr(model, "batch_size_dim", 0))
+            outputs = merge_batch(outputs)
         return {"loss": accum_loss, "outputs": outputs}
 
     def forward_backward_step(

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。

examples/images/vit/vit_benchmark.py

@@ -120,7 +120,7 @@ def main():
                 # run pipeline forward backward
                 batch = iter([batch])
                 outputs = booster.execute_pipeline(
-                    batch, model, criterion, optimizer, return_loss=True, return_outputs=True
+                    batch, model, criterion, optimizer, return_loss=True
                 )
             else:
                 outputs = model(**batch)

examples/language/bert/finetune.py

@@ -148,7 +148,7 @@ 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_device:

examples/language/gpt/hybridparallelism/finetune.py

@@ -145,7 +145,7 @@ 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:

examples/language/llama2/finetune.py

@@ -271,7 +271,7 @@ def main():
             for step in pbar:
                 if use_pipeline:
                     outputs = booster.execute_pipeline(
-                        dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True
+                        dataloader_iter, model, _criterion, optimizer, return_loss=True
                     )
                     loss = outputs["loss"]
                 else:

examples/language/llama2/pretrain.py

@@ -185,7 +185,7 @@ def main():
             microbatch_size=1,
             enable_jit_fused=False,
             zero_stage=0,
-            precision="fp32",
+            precision=args.mixed_precision,
             initial_scale=1,
         )
     else:
@@ -286,7 +286,7 @@ def main():
             for step in pbar:
                 if use_pipeline:
                     outputs = booster.execute_pipeline(
-                        dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True
+                        dataloader_iter, model, _criterion, optimizer, return_loss=True
                     )
                     loss = outputs["loss"]
                 else:

examples/language/openmoe/benchmark/benchmark_cai.py

@@ -270,7 +270,6 @@ def main():
                     lambda x, y: x.loss,
                     optimizer,
                     return_loss=True,
-                    return_outputs=True,
                 )
                 # Backward and optimize
                 if is_pp_last_stage:

examples/language/openmoe/train.py

@@ -340,7 +340,6 @@ def main():
                         lambda x, y: x.loss,
                         optimizer,
                         return_loss=True,
-                        return_outputs=True,
                     )
                     # Backward and optimize
                     if is_pp_last_stage:

examples/language/opt/opt_train_demo.py

@@ -42,7 +42,7 @@ def train_epoch(epoch, model, optimizer, _criterion, lr_scheduler, dataloader, b
         for _ in pbar:
             if use_pipeline:
                 outputs = booster.execute_pipeline(
-                    dataloader, model, _criterion, optimizer, return_loss=True, return_outputs=True
+                    dataloader, model, _criterion, optimizer, return_loss=True
                 )
                 # Backward and optimize
                 if is_pp_last_stage:

tests/test_booster/test_plugin/test_3d_plugin.py

@@ -74,7 +74,7 @@ def run_fn(init_method, model_fn, data_gen_fn, output_transform_fn) -> Optional[
             loss = criterion(outputs[output_key])
             return loss
 
-        booster.execute_pipeline(data_iter, model, _criterion, optimizer, return_loss=True, return_outputs=False)
+        booster.execute_pipeline(data_iter, model, _criterion, optimizer, return_loss=True)
         optimizer.step()
 
     except Exception as e: