ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Update rl_replacements.py

Browse source
This commit is contained in:
Daniel Han 2025-06-25 19:51:43 -07:00
commit 580842de0b
1 changed files with 2 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

3
unsloth/models/rl_replacements.py
View file

@ -324,7 +324,8 @@ def grpo_trainer_compute_loss(function_name, function):
print("logits_to_keep Unsloth 320", logits_to_keep)
per_token_logps = self._get_per_token_logps(model, input_ids, attention_mask, logits_to_keep)
print("per_token_logps Unsloth 320", per_token_logps.shape)
if per_token_logps is not None:
print("per_token_logps Unsloth 320", per_token_logps.shape)
# Compute the KL divergence between the model and the reference model
# _prepare_inputs doesn't return reference log probs anymore. We need to calculate it ourselves.