[trainer] self.model_wrapped + _model_unwrap (#9390)

* model wrapped + model_unwrap

* cleanup

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* style

* deprecation warning

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/trainer.py

@@ -162,6 +162,14 @@ if is_fairscale_available():
 logger = logging.get_logger(__name__)
 
 
+def _model_unwrap(model: nn.Module) -> nn.Module:
+    # since there could be multiple levels of wrapping, unwrap recursively
+    if hasattr(model, "module"):
+        return _model_unwrap(model.module)
+    else:
+        return model
+
+
 class Trainer:
     """
     Trainer is a simple but feature-complete training and eval loop for PyTorch, optimized for 🤗 Transformers.
@@ -212,6 +220,16 @@ class Trainer:
             containing the optimizer and the scheduler to use. Will default to an instance of
             :class:`~transformers.AdamW` on your model and a scheduler given by
             :func:`~transformers.get_linear_schedule_with_warmup` controlled by :obj:`args`.
+
+    Important accessors:
+
+        ``self.model`` - always points to the core model. If using a transformers model, it will be a
+        :class:`PreTrainedModel` subclass.
+
+        ``self.model_wrapped`` - always points to the most external model in case one or more other modules wrap the
+        original model. This is the model that should be used for the forward pass. For example, under ``DeepSpeed``,
+        the inner model is wrapped in ``DeepSpeed`` and then again in ``DistributedDataParallel``. If the inner model
+        hasn't been wrapped, then ``self.model_wrapped`` is the same as ``self.model``.
     """
 
     def __init__(
@@ -234,30 +252,37 @@ class Trainer:
         self.args = args
         # Seed must be set before instantiating the model when using model
         set_seed(self.args.seed)
-        assert (
-            model is not None or model_init is not None
-        ), "You must provide a model to use `Trainer`, either by using the `model` argument or the `model_init` argument."
-        self.model_init = model_init
         self.hp_name = None
-        if model is None and model_init is not None:
-            model = self.call_model_init()
-
-        if self.args.model_parallel and not model.is_parallelizable:
-            raise ValueError(
-                f"{model.__class__.__name__} implementation currently doesn't support model parallelism, therefore --model_parallel cl arg cannot be used"
-            )
 
-        # Model parallel
-        if model is not None and not self.args.model_parallel:
-            model = model.to(args.device)
+        if model is None:
+            if model_init is not None:
+                self.model_init = model_init
+                model = self.call_model_init()
+            else:
+                raise RuntimeError("`Trainer` requires either a `model` or `model_init` argument")
+        else:
+            if model_init is not None:
+                warnings.warn(
+                    "`Trainer` requires either a `model` or `model_init` argument, but not both. "
+                    "`model_init` will overwrite your model when calling the `train` method. This will become a fatal error in the next release.",
+                    FutureWarning,
+                )
+            self.model_init = model_init
 
-        self.model = model
         default_collator = default_data_collator if tokenizer is None else DataCollatorWithPadding(tokenizer)
         self.data_collator = data_collator if data_collator is not None else default_collator
         self.train_dataset = train_dataset
         self.eval_dataset = eval_dataset
         self.tokenizer = tokenizer
 
+        # Model parallel
+        if not self.args.model_parallel:
+            model = model.to(args.device)
+
+        # later use `self.model is self.model_wrapped` to check if it's wrapped or not
+        self.model_wrapped = model
+        self.model = model
+
         self.compute_metrics = compute_metrics
         self.optimizer, self.lr_scheduler = optimizers
         if model_init is not None and (self.optimizer is not None or self.lr_scheduler is not None):
@@ -640,9 +665,11 @@ class Trainer:
             set_seed(self.args.seed)
 
             model = self.call_model_init(trial)
-
             if not self.args.model_parallel:
-                self.model = model.to(self.args.device)
+                model = model.to(self.args.device)
+
+            self.model = model
+            self.model_wrapped = model
 
             # Reinitializes optimizer and scheduler
             self.optimizer, self.lr_scheduler = None, None
@@ -681,8 +708,9 @@ class Trainer:
         # Check if saved optimizer or scheduler states exist
         self._load_optimizer_and_scheduler(model_path)
 
+        model = self.model_wrapped
+
         # Mixed precision training with apex (torch < 1.6)
-        model = self.model
         if self.use_apex:
             model, self.optimizer = amp.initialize(model, self.optimizer, opt_level=self.args.fp16_opt_level)
 
@@ -707,6 +735,14 @@ class Trainer:
             # find_unused_parameters breaks checkpointing as per
             # https://github.com/huggingface/transformers/pull/4659#issuecomment-643356021
 
+        # for the rest of this function `model` is the outside model, whether it was wrapped or not
+        if model is not self.model:
+            self.model_wrapped = model
+
+        # important: at this point:
+        # self.model         is the Transformers Model
+        # self.model_wrapped is DDP(Transformers Model), DDP(Deepspeed(Transformers Model)), etc.
+
         # Train!
         if is_torch_tpu_available():
             total_train_batch_size = self.args.train_batch_size * xm.xrt_world_size()
@@ -937,12 +973,10 @@ class Trainer:
             self.control = self.callback_handler.on_save(self.args, self.state, self.control)
 
     def _save_checkpoint(self, model, trial, metrics=None):
-        # In all cases (even distributed/parallel), self.model is always a reference
-        # to the model we want to save.
-        if hasattr(model, "module"):
-            assert model.module is self.model, f"Module {model.module} should be a reference to self.model"
-        else:
-            assert model is self.model, f"Model {model} should be a reference to self.model"
+        # In all cases, including ddp/dp/deepspeed, self.model is always a reference to the model we
+        # want to save.
+        assert _model_unwrap(model) is self.model, "internal model should be a reference to self.model"
+
         # Save model checkpoint
         checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
 
@@ -1630,30 +1664,7 @@ class Trainer:
         Returns:
             :obj:`int`: The number of floating-point operations.
         """
-
-        model = self._actual_model(self.model)
-
-        if hasattr(model, "floating_point_ops"):
-            return model.floating_point_ops(inputs)
-
+        if hasattr(self.model, "floating_point_ops"):
+            return self.model.floating_point_ops(inputs)
         else:
             return 0
-
-    @staticmethod
-    def _actual_model(
-        model: Union[torch.nn.DataParallel, torch.nn.parallel.DistributedDataParallel, torch.nn.modules.Module]
-    ) -> torch.nn.modules.Module:
-        """
-
-        Args:
-            model: (:obj:`Union[torch.nn.DataParallel, torch.nn.parallel.DistributedDataParallel, torch.nn.modules.Module]`):
-                Model object used during training
-
-        Returns:
-            :obj:`torch.nn.modules.Module`: unwrapped module
-        """
-        if isinstance(model, torch.nn.DataParallel) or isinstance(model, torch.nn.parallel.DistributedDataParallel):
-            model = model.module
-        else:
-            model = model
-        return model

tests/test_trainer.py

@@ -53,6 +53,7 @@ if is_torch_available():
         Trainer,
         TrainerState,
     )
+    from transformers.trainer import _model_unwrap
 
 
 PATH_SAMPLE_TEXT = f"{get_tests_dir()}/fixtures/sample_text.txt"
@@ -850,8 +851,8 @@ class TrainerIntegrationTest(unittest.TestCase):
         trainer = get_regression_trainer(learning_rate=0.1)
 
         def assert_flos_extraction(trainer, wrapped_model_to_check):
-            self.assertEqual(trainer.model, trainer._actual_model(wrapped_model_to_check))
-            self.assertGreaterEqual(getattr(trainer._actual_model(wrapped_model_to_check).config, "total_flos", 0), 0)
+            self.assertEqual(trainer.model, _model_unwrap(wrapped_model_to_check))
+            self.assertGreaterEqual(getattr(_model_unwrap(wrapped_model_to_check).config, "total_flos", 0), 0)
 
         # with plain model
         assert_flos_extraction(trainer, trainer.model)