frozenleaves/trl
1
0
Fork 0
mirror of https://github.com/huggingface/trl.git synced 2025-10-20 10:03:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

set token scaling flag

Browse Source
This commit is contained in:
Kashif Rasul
2025-10-18 13:31:30 +02:00
parent 622ee4ac5f
commit 1b9ec9f0cd
1 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
trl/trainer/sft_trainer.py
View File

@ -818,7 +818,7 @@ class SFTTrainer(BaseTrainer):
)
# Loss function
if args.loss_type == "nll":
if args.loss_type == "nll" or args.use_liger_kernel:
pass # use the default loss
elif args.loss_type == "dft":
if compute_loss_func is not None:
@ -1095,9 +1095,10 @@ class SFTTrainer(BaseTrainer):
# If not set, defaults from model config and may warn since cache isn't compatible with gradient checkpointing
inputs["use_cache"] = False
# Request token accuracy from Liger kernel if used
# Request token accuracy from Liger kernel and set token scaling if using DFT loss
if self.args.use_liger_kernel:
inputs["return_token_accuracy"] = True
inputs["use_token_scaling"] = self.args.loss_type == "dft"
(loss, outputs) = super().compute_loss(
model, inputs, return_outputs=True, num_items_in_batch=num_items_in_batch