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/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