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

rl: guard warnings_issued before TRL estimate_tokens write (#4034)

Browse source 
Co-authored-by: Daniel Hanchen <danielhanchen@users.noreply.github.com>
This commit is contained in:
Daniel Han 2026-03-01 00:42:37 -08:00 • committed by GitHub
commit fc63b7ec4d
1 changed files with 12 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

12
unsloth/models/rl.py
View file

@ -810,6 +810,18 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
" os.environ['UNSLOTH_RETURN_LOGITS'] = '1'\n"
)
extra_args += logits_check
warnings_issued_check = (
"if model is not None:\n"
" _warnings_issued = getattr(model, 'warnings_issued', None)\n"
" if _warnings_issued is None:\n"
" model.warnings_issued = {}\n"
" elif not isinstance(_warnings_issued, dict):\n"
" try:\n"
" model.warnings_issued = dict(_warnings_issued)\n"
" except Exception:\n"
" model.warnings_issued = {}\n"
)
extra_args += warnings_issued_check
# Check max_seq_length
if "model" in call_args: