423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-18 12:50:38 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix incorrect loss value for vlms

Browse Source 
Former-commit-id: 30567a1487
This commit is contained in:
hiyouga
2024-10-30 08:56:46 +00:00
parent 1b02915d19
commit 584ce3a105
12 changed files with 48 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/llamafactory/train/dpo/trainer.py
View File

@@ -101,7 +101,7 @@ class CustomDPOTrainer(DPOTrainer):
self.callback_handler.add_callback(PissaConvertCallback)
if finetuning_args.use_badam:
from badam import BAdamCallback, clip_grad_norm_old_version
from badam import BAdamCallback, clip_grad_norm_old_version # type: ignore
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
self.add_callback(BAdamCallback)
@@ -274,7 +274,7 @@ class CustomDPOTrainer(DPOTrainer):
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():
if is_transformers_version_equal_to_4_46() and kwargs.pop("num_items_in_batch", False):
loss /= self.args.gradient_accumulation_steps
return loss

4
src/llamafactory/train/kto/trainer.py
View File

@@ -96,7 +96,7 @@ class CustomKTOTrainer(KTOTrainer):
self.add_callback(SaveProcessorCallback(processor))
if finetuning_args.use_badam:
from badam import BAdamCallback, clip_grad_norm_old_version
from badam import BAdamCallback, clip_grad_norm_old_version # type: ignore
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
self.add_callback(BAdamCallback)
@@ -247,7 +247,7 @@ class CustomKTOTrainer(KTOTrainer):
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():
if is_transformers_version_equal_to_4_46() and kwargs.pop("num_items_in_batch", False):
loss /= self.args.gradient_accumulation_steps
return loss

2
src/llamafactory/train/ppo/trainer.py
View File

@@ -181,7 +181,7 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
self.add_callback(SaveProcessorCallback(processor))
if finetuning_args.use_badam:
from badam import BAdamCallback, clip_grad_norm_old_version
from badam import BAdamCallback, clip_grad_norm_old_version # type: ignore
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
self.add_callback(BAdamCallback)

15
src/llamafactory/train/pt/trainer.py
View File

@@ -19,6 +19,7 @@ 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 PissaConvertCallback, SaveProcessorCallback
from ..trainer_utils import create_custom_optimizer, create_custom_scheduler
@@ -51,7 +52,7 @@ class CustomTrainer(Trainer):
self.add_callback(PissaConvertCallback)
if finetuning_args.use_badam:
from badam import BAdamCallback, clip_grad_norm_old_version
from badam import BAdamCallback, clip_grad_norm_old_version # type: ignore
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
self.add_callback(BAdamCallback)
@@ -68,3 +69,15 @@ class CustomTrainer(Trainer):
) -> "torch.optim.lr_scheduler.LRScheduler":
create_custom_scheduler(self.args, num_training_steps, optimizer)
return super().create_scheduler(num_training_steps, optimizer)
@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, **kwargs)
if is_transformers_version_equal_to_4_46() and not getattr(self, "model_accepts_loss_kwargs", False):
loss /= self.args.gradient_accumulation_steps # other model should not scale the loss
return loss

4
src/llamafactory/train/rm/trainer.py
View File

@@ -60,7 +60,7 @@ class PairwiseTrainer(Trainer):
self.add_callback(PissaConvertCallback)
if finetuning_args.use_badam:
from badam import BAdamCallback, clip_grad_norm_old_version
from badam import BAdamCallback, clip_grad_norm_old_version # type: ignore
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
self.add_callback(BAdamCallback)
@@ -100,7 +100,7 @@ class PairwiseTrainer(Trainer):
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():
if is_transformers_version_equal_to_4_46() and kwargs.pop("num_items_in_batch", False):
loss /= self.args.gradient_accumulation_steps # fixes the loss value for transformers 4.46.0
if return_outputs:

15
src/llamafactory/train/sft/trainer.py
View File

@@ -27,6 +27,7 @@ from typing_extensions import override
from ...extras.constants import IGNORE_INDEX
from ...extras.logging import get_logger
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
@@ -60,7 +61,7 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
self.add_callback(PissaConvertCallback)
if finetuning_args.use_badam:
from badam import BAdamCallback, clip_grad_norm_old_version
from badam import BAdamCallback, clip_grad_norm_old_version # type: ignore
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
self.add_callback(BAdamCallback)
@@ -78,6 +79,18 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
create_custom_scheduler(self.args, num_training_steps, optimizer)
return super().create_scheduler(num_training_steps, optimizer)
@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, **kwargs)
if is_transformers_version_equal_to_4_46() and not getattr(self, "model_accepts_loss_kwargs", False):
loss /= self.args.gradient_accumulation_steps # other model should not scale the loss
return loss
@override
def prediction_step(
self,