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 2025-06-22 02:37:52 -07:00
commit 3fa2f1436a
1 changed files with 5 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

6
unsloth/models/rl_replacements.py
View file

@ -320,12 +320,16 @@ def grpo_trainer_compute_loss(function_name, function):
# _prepare_inputs doesn't return reference log probs anymore. We need to calculate it ourselves.
# https://github.com/huggingface/trl/blob/05bc43e960396581e458195b8388efe6b82cae1f/trl/trainer/grpo_trainer.py#L1328
if self.beta != 0.0:
print("!!!!!!!!!!")
print("!!!!!!!!!!")
with torch.inference_mode(), model.disable_adapter():
ref_per_token_logps = self._get_per_token_logps(model, input_ids, attention_mask, logits_to_keep)
else:
ref_per_token_logps = None
# per_token_kl = torch.exp(ref_per_token_logps - per_token_logps) - (ref_per_token_logps - per_token_logps) - 1
print("!!!!!!!!!!")
print("!!!!!!!!!!")
print(ref_per_token_logps)
# x - x.detach() allows for preserving gradients from x
advantages = inputs["advantages"]
# per_token_loss = torch.exp(per_token_logps - per_token_logps.detach()) * advantages.unsqueeze(1)