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

Merge pull request #3842 from unslothai/fix-vllm-chat-template-sync

Browse source 
Sync chat_template from tokenizer to vLLM
This commit is contained in:
Daniel Han 2026-01-05 04:38:39 -08:00 • committed by GitHub
commit 3e193d08c3
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 13 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

13
unsloth/models/rl.py
View file

@ -694,6 +694,19 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
)
RLTrainer_post += training_check
# Sync chat_template from processing_class to vLLM's tokenizer
# This fixes base models that have custom chat templates applied after loading
if "model" in call_args:
vllm_chat_template_sync = (
"if hasattr(self, 'llm') and self.llm is not None and hasattr(self.llm, 'get_tokenizer'):\n"
" _vllm_tok = self.llm.get_tokenizer()\n"
" _pc = getattr(self, 'processing_class', None) or getattr(self, 'tokenizer', None)\n"
" if _vllm_tok is not None and _pc is not None and getattr(_pc, 'chat_template', None) is not None and getattr(_vllm_tok, 'chat_template', None) is None:\n"
" _vllm_tok.chat_template = _pc.chat_template\n"
"pass\n"
)
RLTrainer_post += vllm_chat_template_sync
# Edit optional metrics
other_metrics_processor = ""
if trainer_file in RL_METRICS_CHANGES: