423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-16 20:00:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix incorrect loss value for vlms

Browse Source
This commit is contained in:
hiyouga
2024-10-30 08:56:46 +00:00
parent 0c22da4f1c
commit 30567a1487
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/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: