fix #5747

Former-commit-id: 26d07de349c98b547cd6a6166ea20616d08ba343

src/llamafactory/extras/packages.py

@@ -79,6 +79,11 @@ def is_transformers_version_greater_than_4_43():
     return _get_package_version("transformers") >= version.parse("4.43.0")
 
 
+@lru_cache
+def is_transformers_version_equal_to_4_46():
+    return _get_package_version("transformers") == version.parse("4.46.0")
+
+
 def is_uvicorn_available():
     return _is_package_available("uvicorn")
 

src/llamafactory/train/dpo/trainer.py

@@ -29,6 +29,7 @@ from trl.trainer import disable_dropout_in_model
 from typing_extensions import override
 
 from ...extras.constants import IGNORE_INDEX
+from ...extras.packages import is_transformers_version_equal_to_4_46
 from ..callbacks import PissaConvertCallback, SaveProcessorCallback
 from ..trainer_utils import create_custom_optimizer, create_custom_scheduler, get_batch_logps
 
@@ -118,6 +119,13 @@ class CustomDPOTrainer(DPOTrainer):
         create_custom_scheduler(self.args, num_training_steps, optimizer)
         return super().create_scheduler(num_training_steps, optimizer)
 
+    @override
+    def get_batch_samples(self, epoch_iterator, num_batches):
+        r"""
+        Replaces the method of KTO Trainer with the one of the standard Trainer.
+        """
+        return Trainer.get_batch_samples(self, epoch_iterator, num_batches)
+
     def odds_ratio_loss(self, chosen_logps: "torch.Tensor", rejected_logps: "torch.Tensor") -> "torch.Tensor":
         r"""
         Computes ORPO's odds ratio (OR) loss for batched log probabilities of the policy model.
@@ -258,3 +266,15 @@ class CustomDPOTrainer(DPOTrainer):
             metrics[f"{prefix}odds_ratio_loss"] = ((losses - sft_loss) / self.beta).detach().mean().cpu()
 
         return losses.mean(), metrics
+
+    @override
+    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
+        r"""
+        Fixes the loss value for transformers 4.46.0.
+        https://github.com/huggingface/transformers/blob/v4.46.0/src/transformers/trainer.py#L3605
+        """
+        loss = super().compute_loss(model, inputs, return_outputs)
+        if kwargs.pop("num_items_in_batch", False) and is_transformers_version_equal_to_4_46():
+            loss /= self.args.gradient_accumulation_steps
+
+        return loss

src/llamafactory/train/kto/trainer.py

@@ -28,6 +28,7 @@ from trl.trainer import disable_dropout_in_model
 from typing_extensions import override
 
 from ...extras.constants import IGNORE_INDEX
+from ...extras.packages import is_transformers_version_equal_to_4_46
 from ..callbacks import SaveProcessorCallback
 from ..trainer_utils import create_custom_optimizer, create_custom_scheduler, get_batch_logps
 
@@ -120,6 +121,13 @@ class CustomKTOTrainer(KTOTrainer):
         """
         return Trainer._get_train_sampler(self)
 
+    @override
+    def get_batch_samples(self, epoch_iterator, num_batches):
+        r"""
+        Replaces the method of KTO Trainer with the one of the standard Trainer.
+        """
+        return Trainer.get_batch_samples(self, epoch_iterator, num_batches)
+
     @override
     def forward(
         self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"], prefix: Literal["", "kl_"] = ""
@@ -231,3 +239,15 @@ class CustomKTOTrainer(KTOTrainer):
         metrics["kl"] = kl.item()
 
         return losses, metrics
+
+    @override
+    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
+        r"""
+        Fixes the loss value for transformers 4.46.0.
+        https://github.com/huggingface/transformers/blob/v4.46.0/src/transformers/trainer.py#L3605
+        """
+        loss = super().compute_loss(model, inputs, return_outputs)
+        if kwargs.pop("num_items_in_batch", False) and is_transformers_version_equal_to_4_46():
+            loss /= self.args.gradient_accumulation_steps
+
+        return loss

src/llamafactory/train/rm/trainer.py

@@ -25,6 +25,7 @@ from transformers import Trainer
 from typing_extensions import override
 
 from ...extras.logging import get_logger
+from ...extras.packages import is_transformers_version_equal_to_4_46
 from ..callbacks import FixValueHeadModelCallback, PissaConvertCallback, SaveProcessorCallback
 from ..trainer_utils import create_custom_optimizer, create_custom_scheduler
 
@@ -79,7 +80,7 @@ class PairwiseTrainer(Trainer):
 
     @override
     def compute_loss(
-        self, model: "PreTrainedModel", inputs: Dict[str, "torch.Tensor"], return_outputs: bool = False
+        self, model: "PreTrainedModel", inputs: Dict[str, "torch.Tensor"], return_outputs: bool = False, **kwargs
     ) -> Union["torch.Tensor", Tuple["torch.Tensor", List["torch.Tensor"]]]:
         r"""
         Computes pairwise loss. The first n examples are chosen and the last n examples are rejected.
@@ -98,6 +99,10 @@ class PairwiseTrainer(Trainer):
         chosen_scores, rejected_scores = chosen_scores.squeeze(), rejected_scores.squeeze()
 
         loss = -torch.nn.functional.logsigmoid(chosen_scores.float() - rejected_scores.float()).mean()
+
+        if kwargs.pop("num_items_in_batch", False) and is_transformers_version_equal_to_4_46():
+            loss /= self.args.gradient_accumulation_steps  # fixes the loss value for transformers 4.46.0
+
         if return_outputs:
             return loss, (loss, chosen_scores, rejected_scores)
         else: