Yell at the user if zero-3 init wasn't performed, but expected to have been done (#32299)

* Test this zach

* Test for improper init w/o zero3

* Move back

* Apply suggestions from code review

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Get rid of stars in warning

* Make private

* Make clear

---------

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

src/transformers/modeling_utils.py

@@ -1470,9 +1470,13 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             # and memory copying it on CPU or each GPU first
             with deepspeed.zero.Init(config_dict_or_path=deepspeed_config()):
                 model = cls(config, **kwargs)
+
         else:
             model = cls(config, **kwargs)
 
+        # Flag for if we init with `zero3`, add an attr to the model so we can check downstream for issues
+        model._transformers_zero3_init_used = is_deepspeed_zero3_enabled()
+
         # restore default dtype if it was modified
         if dtype_orig is not None:
             torch.set_default_dtype(dtype_orig)
@@ -3802,6 +3806,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             # Let's make sure we don't run the init function of buffer modules
             model = cls(config, *model_args, **model_kwargs)
 
+        # If we init with `zero3`, add an attr to the model so we can check downstream for issues
+        model._transformers_zero3_init_used = is_deepspeed_zero3_enabled() and not is_quantized
+
         # make sure we use the model's config since the __init__ call might have copied it
         config = model.config
 

src/transformers/trainer.py

@@ -100,6 +100,7 @@ from .trainer_pt_utils import (
     get_model_param_count,
     get_module_class_from_name,
     get_parameter_names,
+    is_deepspeed_zero3_enabled,
     nested_concat,
     nested_detach,
     nested_numpify,
@@ -435,6 +436,15 @@ class Trainer:
                 )
             self.model_init = model_init
 
+        # Will reach this branch if the user has
+        # 1. Used `.from_pretrained` or `.from_config` to initialize their model
+        # 2. Did not configure Zero-3 via `TrainingArguments` or `accelerate launch` beforehand
+        # New models init such as `MyModel()` will not hit this step
+        if is_deepspeed_zero3_enabled() and not getattr(model, "_transformers_zero3_init_used", True):
+            raise ValueError(
+                "Model was not initialized with `Zero-3` despite being configured for DeepSpeed Zero-3. Please re-initialize your model via `Model.from_pretrained(...)` or `Model.from_config(...)` after creating your `TrainingArguments`!"
+            )
+
         if model.__class__.__name__ in MODEL_MAPPING_NAMES:
             raise ValueError(
                 f"The model you have picked ({model.__class__.__name__}) cannot be used as is for training: it only "

tests/deepspeed/test_deepspeed.py

@@ -709,6 +709,31 @@ class TrainerIntegrationDeepSpeed(TrainerIntegrationDeepSpeedWithCustomConfig, T
         # Relative difference. See the note above how to get identical loss on a small bs
         self.assertTrue((no_grad_accum_loss - yes_grad_accum_loss) / (no_grad_accum_loss + 1e-15) <= 1e-3)
 
+    def test_missed_zero3_init(self):
+        from transformers import Trainer  # noqa
+
+        with mockenv_context(**self.dist_env_1_gpu):
+            model = AutoModel.from_pretrained(T5_TINY)
+            training_args = TrainingArguments(
+                output_dir="./test_missed_zero3_init",
+                deepspeed=self.get_config_dict(ZERO3),
+            )
+            with self.assertRaises(
+                ValueError, msg="Model was not initialized with `Zero-3` despite being configured."
+            ):
+                _ = Trainer(
+                    model=model,
+                    args=training_args,
+                )
+            # Now do it properly, triggered from our `TrainingArguments` earlier
+            model = AutoModel.from_pretrained(T5_TINY)
+            trainer = Trainer(
+                model=model,
+                args=training_args,
+            )
+            assert trainer.is_deepspeed_enabled
+            assert model._transformers_zero3_init_used
+
     def check_saved_checkpoints_deepspeed(self, output_dir, freq, total, stage, dtype):
         # adapted from TrainerIntegrationCommon.check_saved_checkpoints
         file_list = [SAFE_WEIGHTS_NAME, "training_args.bin", "trainer_state.json", "config.json"]