Fix GA loss bugs and add unit test (#35121)

* fix GA bugs and add unit test

* narrow down model loss unit test diff gap

* format code to make ruff happy

* send num_items_in_batch argument to decoder

* fix GA loss bug in BertLMHeadModel

* use TinyStories-33M to narrow down diff gap

* fotmat code

* missing .config

* avoid add extra args

---------

Co-authored-by: kangsheng <kangsheng@meituan.com>

src/transformers/models/bert/modeling_bert.py

@@ -1325,6 +1325,7 @@ class BertLMHeadModel(BertPreTrainedModel, GenerationMixin):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **loss_kwargs,
     ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
         r"""
         encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -1375,11 +1376,7 @@ class BertLMHeadModel(BertPreTrainedModel, GenerationMixin):
 
         lm_loss = None
         if labels is not None:
-            # we are doing next-token prediction; shift prediction scores and input ids by one
+            lm_loss = self.loss_function(prediction_scores, labels, self.config.vocab_size, **loss_kwargs)
-            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
-            labels = labels[:, 1:].contiguous()
-            loss_fct = CrossEntropyLoss()
-            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
 
         if not return_dict:
             output = (prediction_scores,) + outputs[2:]

src/transformers/models/speech_encoder_decoder/modeling_speech_encoder_decoder.py

@@ -491,6 +491,8 @@ class SpeechEncoderDecoderModel(PreTrainedModel, GenerationMixin):
         kwargs_decoder = {
             argument[len("decoder_") :]: value for argument, value in kwargs.items() if argument.startswith("decoder_")
         }
+        if "num_items_in_batch" in kwargs_encoder:
+            kwargs_decoder["num_items_in_batch"] = kwargs_encoder.pop("num_items_in_batch", None)
 
         if encoder_outputs is None:
             if inputs is None:

src/transformers/trainer.py

@@ -3649,9 +3649,6 @@ class Trainer:
             return loss_mb.reduce_mean().detach().to(self.args.device)
 
         with self.compute_loss_context_manager():
-            if self.model_accepts_loss_kwargs:
-                loss = self.compute_loss(model, inputs)
-            else:
             loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
 
         del inputs
@@ -5132,10 +5129,6 @@ class Trainer:
             except StopIteration:
                 break
 
-        # Keep default behavior the same
-        if not self.model_accepts_loss_kwargs:
-            return batch_samples, None
-
         if len(batch_samples) > 0 and "labels" in batch_samples[0]:
             # For now we don't support object detection
             try:

tests/trainer/test_trainer.py

@@ -750,11 +750,102 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
         self.check_trained_model(trainer.model, alternate_seed=True)
 
     @slow
-    def test_gradient_accumulation_loss_alignment(self):
+    def test_gradient_accumulation_loss_alignment_with_model_loss(self):
         set_seed(42)
         import datasets
 
-        model_name = "distilgpt2"
+        model_name = "nickypro/tinyllama-110M"
+        dataset_name = "wikitext"
+        dataset_config = "wikitext-2-raw-v1"
+        dataset = datasets.load_dataset(dataset_name, dataset_config, split="train[:500]")
+        dataset = dataset.train_test_split(test_size=0.2)
+        tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+        tokenizer.pad_token = tokenizer.eos_token
+
+        def tokenize_function(examples):
+            return tokenizer(examples["text"], max_length=128, padding="max_length", truncation=True)
+
+        tokenized_dataset = dataset.map(tokenize_function, batched=True, remove_columns=dataset["train"].column_names)
+
+        data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
+
+        model = AutoModelForCausalLM.from_pretrained(model_name)
+
+        base_loss_callback = StoreLossCallback()
+
+        args_kwargs = {
+            "report_to": "none",
+            "logging_steps": 1,
+            "max_steps": 20,
+            "learning_rate": 3e-4,
+            "disable_tqdm": True,
+        }
+
+        args = TrainingArguments(
+            "./generation",
+            **args_kwargs,
+        )
+        trainer = Trainer(
+            model,
+            args,
+            train_dataset=tokenized_dataset["train"],
+            callbacks=[base_loss_callback],
+            data_collator=data_collator,
+        )
+        assert trainer.model_accepts_loss_kwargs
+        trainer.train()
+
+        grad_accum_loss_callback = StoreLossCallback()
+        args = TrainingArguments(
+            "./generation",
+            **args_kwargs,
+            gradient_accumulation_steps=2,
+            per_device_train_batch_size=4,
+        )
+        set_seed(42)
+        model = AutoModelForCausalLM.from_pretrained(model_name)
+        trainer = Trainer(
+            model,
+            args,
+            train_dataset=tokenized_dataset["train"],
+            callbacks=[grad_accum_loss_callback],
+            data_collator=data_collator,
+        )
+        trainer.train()
+
+        set_seed(42)
+        model = AutoModelForCausalLM.from_pretrained(model_name)
+        broken_loss_callback = StoreLossCallback()
+        trainer = Trainer(
+            model,
+            args,
+            train_dataset=tokenized_dataset["train"],
+            callbacks=[broken_loss_callback],
+            data_collator=data_collator,
+        )
+        # disable model_accepts_loss_kwargs
+        trainer.model_accepts_loss_kwargs = False
+        trainer.train()
+
+        # Calculate the difference between the base loss and the grad_accum loss
+        diff_truth = [
+            abs(base - grad) for base, grad in zip(base_loss_callback.losses, grad_accum_loss_callback.losses)
+        ]
+        diff_broken = [abs(base - grad) for base, grad in zip(base_loss_callback.losses, broken_loss_callback.losses)]
+
+        # all diff truth should be quite close
+        self.assertLess(max(diff_truth), 0.01, f"Difference {max(diff_truth)} is not within 0.01")
+
+        # max diff broken should be very off
+        self.assertGreater(max(diff_broken), 3, f"Difference {max(diff_broken)} is not greater than 3")
+
+    @slow
+    def test_gradient_accumulation_loss_alignment_with_loss_func(self):
+        set_seed(42)
+        import datasets
+
+        model_name = "roneneldan/TinyStories-33M"
         dataset_name = "wikitext"
         dataset_config = "wikitext-2-raw-v1"
         dataset = datasets.load_dataset(dataset_name, dataset_config, split="train[:500]")
@@ -836,15 +927,16 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
         trainer.train()
 
         # Calculate the difference between the base loss and the grad_accum loss
-        diff_truth = [base - grad for base, grad in zip(base_loss_callback.losses, grad_accum_loss_callback.losses)]
+        diff_truth = [
-        diff_broken = [base - grad for base, grad in zip(base_loss_callback.losses, broken_loss_callback.losses)]
+            abs(base - grad) for base, grad in zip(base_loss_callback.losses, grad_accum_loss_callback.losses)
-        # These should be quite close
+        ]
-        for diff in diff_truth:
+        diff_broken = [abs(base - grad) for base, grad in zip(base_loss_callback.losses, broken_loss_callback.losses)]
-            self.assertLess(abs(diff), 0.1, f"Difference {diff} is not within 0.1")
 
-        # These should be very off
+        # all diff truth should be quite close
-        for diff in diff_broken:
+        self.assertLess(max(diff_truth), 0.01, f"Difference {max(diff_truth)} is not within 0.01")
-            self.assertGreater(abs(diff), 0.1, f"Difference {diff} is not greater than 0.1")
+
+        # max diff broken should be very off
+        self.assertGreater(max(diff_broken), 3, f"Difference {max(diff_broken)} is not greater than 3")
 
     def test_gradient_accumulation(self):
         # Training with half the batch size but accumulation steps as 2 should give the same training losses.