support disable shuffling

Former-commit-id: 9d8c35fd6b838ede0bd6827c6c6121f2cba2b11b

src/llamafactory/hparams/finetuning_args.py

@@ -342,6 +342,10 @@ class FinetuningArguments(FreezeArguments, LoraArguments, RLHFArguments, GaloreA
         default=False,
         metadata={"help": "Whether or not to compute the token-level accuracy at evaluation."},
     )
+    disable_shuffling: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to disable the shuffling of the training set."},
+    )
     plot_loss: bool = field(
         default=False,
         metadata={"help": "Whether or not to save the training loss curves."},

src/llamafactory/train/dpo/trainer.py

@@ -19,7 +19,7 @@ import warnings
 from collections import defaultdict
 from contextlib import nullcontext
 from types import MethodType
-from typing import TYPE_CHECKING, Dict, Literal, Optional, Tuple, Union
+from typing import TYPE_CHECKING, Dict, List, Literal, Optional, Tuple, Union
 
 import torch
 import torch.nn.functional as F
@@ -119,6 +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_train_sampler(self) -> Optional["torch.utils.data.Sampler"]:
+        if self.finetuning_args.disable_shuffling:
+            return torch.utils.data.SequentialSampler(self.train_dataset)
+
+        return super()._get_train_sampler()
+
     @override
     def get_batch_samples(self, epoch_iterator, num_batches):
         r"""
@@ -266,7 +273,9 @@ class CustomDPOTrainer(DPOTrainer):
         return losses.mean(), metrics
 
     @override
-    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
+    def compute_loss(
+        self, model: "PreTrainedModel", inputs: Dict[str, "torch.Tensor"], return_outputs: bool = False, **kwargs
+    ) -> Union["torch.Tensor", Tuple["torch.Tensor", List["torch.Tensor"]]]:
         r"""
         Fixes the loss value for transformers 4.46.0.
         https://github.com/huggingface/transformers/blob/v4.46.0/src/transformers/trainer.py#L3605

src/llamafactory/train/kto/trainer.py

@@ -19,7 +19,7 @@ import warnings
 from collections import defaultdict
 from contextlib import nullcontext
 from types import MethodType
-from typing import TYPE_CHECKING, Dict, Literal, Optional, Tuple, Union
+from typing import TYPE_CHECKING, Dict, List, Literal, Optional, Tuple, Union
 
 import torch
 from transformers import Trainer
@@ -119,6 +119,9 @@ class CustomKTOTrainer(KTOTrainer):
         r"""
         Replaces the sequential sampler of KTO Trainer created by trl with the random sampler.
         """
+        if self.finetuning_args.disable_shuffling:
+            return torch.utils.data.SequentialSampler(self.train_dataset)
+
         return Trainer._get_train_sampler(self)
 
     @override
@@ -245,7 +248,9 @@ class CustomKTOTrainer(KTOTrainer):
         return losses, metrics
 
     @override
-    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
+    def compute_loss(
+        self, model: "PreTrainedModel", inputs: Dict[str, "torch.Tensor"], return_outputs: bool = False, **kwargs
+    ) -> Union["torch.Tensor", Tuple["torch.Tensor", List["torch.Tensor"]]]:
         r"""
         Fixes the loss value for transformers 4.46.0.
         https://github.com/huggingface/transformers/blob/v4.46.0/src/transformers/trainer.py#L3605

src/llamafactory/train/pt/trainer.py

@@ -13,8 +13,9 @@
 # limitations under the License.
 
 from types import MethodType
-from typing import TYPE_CHECKING, Optional
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
 
+import torch
 from transformers import Trainer
 from typing_extensions import override
 
@@ -24,8 +25,7 @@ from ..trainer_utils import create_custom_optimizer, create_custom_scheduler
 
 
 if TYPE_CHECKING:
-    import torch
-    from transformers import ProcessorMixin
+    from transformers import PreTrainedModel, ProcessorMixin
 
     from ...hparams import FinetuningArguments
 
@@ -70,7 +70,16 @@ class CustomTrainer(Trainer):
         return super().create_scheduler(num_training_steps, optimizer)
 
     @override
-    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
+    def _get_train_sampler(self) -> Optional["torch.utils.data.Sampler"]:
+        if self.finetuning_args.disable_shuffling:
+            return torch.utils.data.SequentialSampler(self.train_dataset)
+
+        return super()._get_train_sampler()
+
+    @override
+    def compute_loss(
+        self, model: "PreTrainedModel", inputs: Dict[str, "torch.Tensor"], return_outputs: bool = False, **kwargs
+    ) -> Union["torch.Tensor", Tuple["torch.Tensor", List["torch.Tensor"]]]:
         r"""
         Fixes the loss value for transformers 4.46.0.
         https://github.com/huggingface/transformers/blob/v4.46.0/src/transformers/trainer.py#L3605

src/llamafactory/train/rm/trainer.py

@@ -81,6 +81,13 @@ class PairwiseTrainer(Trainer):
         create_custom_scheduler(self.args, num_training_steps, optimizer)
         return super().create_scheduler(num_training_steps, optimizer)
 
+    @override
+    def _get_train_sampler(self) -> Optional["torch.utils.data.Sampler"]:
+        if self.finetuning_args.disable_shuffling:
+            return torch.utils.data.SequentialSampler(self.train_dataset)
+
+        return super()._get_train_sampler()
+
     @override
     def compute_loss(
         self, model: "PreTrainedModel", inputs: Dict[str, "torch.Tensor"], return_outputs: bool = False, **kwargs

src/llamafactory/train/sft/trainer.py

@@ -34,7 +34,7 @@ from ..trainer_utils import create_custom_optimizer, create_custom_scheduler
 
 if TYPE_CHECKING:
     from torch.utils.data import Dataset
-    from transformers import PreTrainedTokenizer, ProcessorMixin
+    from transformers import PreTrainedModel, PreTrainedTokenizer, ProcessorMixin
     from transformers.trainer import PredictionOutput
 
     from ...hparams import FinetuningArguments
@@ -85,7 +85,16 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
         return super().create_scheduler(num_training_steps, optimizer)
 
     @override
-    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
+    def _get_train_sampler(self) -> Optional["torch.utils.data.Sampler"]:
+        if self.finetuning_args.disable_shuffling:
+            return torch.utils.data.SequentialSampler(self.train_dataset)
+
+        return super()._get_train_sampler()
+
+    @override
+    def compute_loss(
+        self, model: "PreTrainedModel", inputs: Dict[str, "torch.Tensor"], return_outputs: bool = False, **kwargs
+    ) -> Union["torch.Tensor", Tuple["torch.Tensor", List["torch.Tensor"]]]:
         r"""
         Fixes the loss value for transformers 4.46.0.
         https://github.com/huggingface/transformers/blob/v4.46.0/src/transformers/trainer.py#L3605