ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[breaking] support transformers 4.48 (#6628)

Browse Source 
Former-commit-id: f154ab175c513a4d7bb866bf2cffc34b77b50508
This commit is contained in:
hoshi-hiyouga
2025-01-31 01:36:33 +08:00
committed by GitHub
parent e71737351f
commit 222423bcef
17 changed files with 53 additions and 105 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
src/llamafactory/train/dpo/trainer.py
View File

@@ -29,7 +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, is_transformers_version_greater_than
from ...extras.packages import is_transformers_version_greater_than
from ..callbacks import SaveProcessorCallback
from ..trainer_utils import create_custom_optimizer, create_custom_scheduler, get_batch_logps, nested_detach
@@ -282,19 +282,12 @@ class CustomDPOTrainer(DPOTrainer):
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. See https://github.com/huggingface/transformers/pull/35438 for details.
Subclass and override to accept extra kwargs.
"""
loss = super().compute_loss(model, inputs, return_outputs)
if is_transformers_version_equal_to_4_46() and kwargs.get("num_items_in_batch"):
if return_outputs:
loss = (loss[0] / self.args.gradient_accumulation_steps, *loss[1:])
else:
loss = loss / self.args.gradient_accumulation_steps
return loss
return super().compute_loss(model, inputs, return_outputs)
@override
def log(self, logs: Dict[str, float]) -> None:
def log(self, logs: Dict[str, float], *args, **kwargs) -> None:
r"""
Log `logs` on the various objects watching training, including stored metrics.
"""
@@ -318,4 +311,4 @@ class CustomDPOTrainer(DPOTrainer):
if not key.startswith("dummy_"):
logs[key] = metric
return Trainer.log(self, logs)
return Trainer.log(self, logs, *args, **kwargs)

17
src/llamafactory/train/kto/trainer.py
View File

@@ -28,7 +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, is_transformers_version_greater_than
from ...extras.packages import is_transformers_version_greater_than
from ..callbacks import SaveProcessorCallback
from ..trainer_utils import create_custom_optimizer, create_custom_scheduler, get_batch_logps, nested_detach
@@ -256,19 +256,12 @@ class CustomKTOTrainer(KTOTrainer):
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. See https://github.com/huggingface/transformers/pull/35438 for details.
Subclass and override to accept extra kwargs.
"""
loss = super().compute_loss(model, inputs, return_outputs)
if is_transformers_version_equal_to_4_46() and kwargs.get("num_items_in_batch"):
if return_outputs:
loss = (loss[0] / self.args.gradient_accumulation_steps, *loss[1:])
else:
loss = loss / self.args.gradient_accumulation_steps
return loss
return super().compute_loss(model, inputs, return_outputs)
@override
def log(self, logs: Dict[str, float]) -> None:
def log(self, logs: Dict[str, float], *args, **kwargs) -> None:
r"""
Log `logs` on the various objects watching training, including stored metrics.
"""
@@ -304,4 +297,4 @@ class CustomKTOTrainer(KTOTrainer):
if not key.startswith("dummy_"):
logs[key] = metric
return Trainer.log(self, logs)
return Trainer.log(self, logs, *args, **kwargs)

22
src/llamafactory/train/pt/trainer.py
View File

@@ -13,7 +13,7 @@
# limitations under the License.
from types import MethodType
from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
from typing import TYPE_CHECKING, Optional
import torch
from transformers import Trainer
@@ -25,7 +25,7 @@ from ..trainer_utils import create_custom_optimizer, create_custom_scheduler
if TYPE_CHECKING:
from transformers import PreTrainedModel, ProcessorMixin
from transformers import ProcessorMixin
from ...hparams import FinetuningArguments
@@ -72,21 +72,3 @@ class CustomTrainer(Trainer):
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. See https://github.com/huggingface/transformers/pull/35438 for details.
It should be removed after https://github.com/huggingface/transformers/pull/35651 is merged.
"""
loss = super().compute_loss(model, inputs, return_outputs, **kwargs)
if kwargs.get("num_items_in_batch") and not getattr(self, "model_accepts_loss_kwargs", False):
if return_outputs:
loss = (loss[0] / self.args.gradient_accumulation_steps, *loss[1:])
else:
loss = loss / self.args.gradient_accumulation_steps
return loss

6
src/llamafactory/train/rm/trainer.py
View File

@@ -25,7 +25,7 @@ from transformers import Trainer
from typing_extensions import override
from ...extras import logging
from ...extras.packages import is_transformers_version_equal_to_4_46, is_transformers_version_greater_than
from ...extras.packages import is_transformers_version_greater_than
from ..callbacks import FixValueHeadModelCallback, SaveProcessorCallback
from ..trainer_utils import create_custom_optimizer, create_custom_scheduler
@@ -107,10 +107,6 @@ 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 is_transformers_version_equal_to_4_46() and kwargs.get("num_items_in_batch"):
loss /= self.args.gradient_accumulation_steps # fixes the loss value for transformers 4.46.0-4.46.1
if return_outputs:
return loss, (loss, chosen_scores, rejected_scores)
else:

20
src/llamafactory/train/sft/trainer.py
View File

@@ -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 PreTrainedModel, PreTrainedTokenizer, ProcessorMixin
from transformers import PreTrainedTokenizer, ProcessorMixin
from transformers.trainer import PredictionOutput
from ...hparams import FinetuningArguments
@@ -88,24 +88,6 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
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. See https://github.com/huggingface/transformers/pull/35438 for details.
It should be removed after https://github.com/huggingface/transformers/pull/35651 is merged.
"""
loss = super().compute_loss(model, inputs, return_outputs, **kwargs)
if kwargs.get("num_items_in_batch") and not getattr(self, "model_accepts_loss_kwargs", False):
if return_outputs:
loss = (loss[0] / self.args.gradient_accumulation_steps, *loss[1:])
else:
loss = loss / self.args.gradient_accumulation_steps
return loss
@override
def prediction_step(
self,