Update rl_replacements.py

This commit is contained in:
Daniel Han 2025-02-17 00:43:11 -08:00
commit a9a4b5e436

View file

@ -233,11 +233,11 @@ def grpo_trainer_compute_loss(function_name, function):
loss, completion_length, mean_kl = grpo_compute_loss(
ref_per_token_logps, per_token_logps, input_ids, completion_mask, self.beta, advantages,
)
from unsloth_zoo.rl_replacements import RL_REPLACEMENTS
RL_REPLACEMENTS["data"] = (
ref_per_token_logps.detach(), per_token_logps.detach(), _input_ids, completion_mask, self.beta, advantages,
loss.detach(), completion_length, mean_kl, completion_ids, _logits_to_keep,
)
from unsloth_zoo.rl_replacements import RL_REPLACEMENTS
if "count" in RL_REPLACEMENTS:
RL_REPLACEMENTS["count"] += 1
if RL_REPLACEMENTS["count"] == 10: raise