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

fix #4120

Browse Source 
Former-commit-id: 2a44da678a5e360a9c0f9056397ac9e801329321
This commit is contained in:
hiyouga
2024-06-07 04:18:05 +08:00
parent 8c4c2e580c
commit d0edcde4ea
7 changed files with 47 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/llamafactory/train/dpo/trainer.py
View File

@@ -10,7 +10,7 @@ from trl import DPOTrainer
from trl.trainer import disable_dropout_in_model
from ...extras.constants import IGNORE_INDEX
from ..trainer_utils import create_custom_optimzer, create_custom_scheduler, get_ref_context
from ..trainer_utils import create_custom_optimzer, create_custom_scheduler, get_batch_logps, get_ref_context
if TYPE_CHECKING:
@@ -155,12 +155,7 @@ class CustomDPOTrainer(DPOTrainer):
all_logits: "torch.Tensor" = model(**batch, return_dict=True, use_cache=False).logits.to(torch.float32)
all_logps, valid_length = self.get_batch_logps(
logits=all_logits,
labels=batch["labels"],
is_encoder_decoder=self.is_encoder_decoder,
label_pad_token_id=self.label_pad_token_id,
)
all_logps, valid_length = get_batch_logps(logits=all_logits, labels=batch["labels"])
if self.loss_type in ["ipo", "orpo", "simpo"]:
all_logps = all_logps / valid_length