[Bugfix][Scoring] Warn when serving original Qwen3 reranker without chat template (#56017)

Signed-off-by: 김재억 <gadian88@gmail.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
김재억
2026-09-11 12:52:45 +00:00
committed by GitHub
co-authored by Claude Opus 5
parent 89dbb26445
commit 912dfb3758
2 changed files with 168 additions and 0 deletions
@@ -7,6 +7,7 @@ from typing import Any, TypeAlias, cast
import torch.nn.functional as F
from vllm import PoolingParams, PoolingRequestOutput, TokensPrompt
from vllm.logger import init_logger
from vllm.renderers import TokenizeParams
from vllm.renderers.hf import safe_apply_chat_template
from vllm.renderers.inputs.preprocess import (
@@ -47,6 +48,8 @@ from .utils import (
validate_score_input,
)
logger = init_logger(__name__)
ScoringServeContext: TypeAlias = PoolingServeContext[ScoringRequest]
@@ -425,6 +428,24 @@ class CrossEncoderIOProcessor(ScoringIOProcessor):
self.model = model if self.supports_score_template else None
self.use_sep_token = self.model_config.use_sep_token
if (
getattr(self.model_config.hf_config, "is_original_qwen3_reranker", False)
and self.chat_template is None
):
suggested_template = (
"examples/pooling/score/template/qwen3_vl_reranker.jinja"
if self.model_config.is_multimodal_model
else "examples/pooling/score/template/qwen3_reranker.jinja"
)
logger.warning(
"Serving an original Qwen3 reranker (%s) without a "
"--chat-template. The model was trained with a specific prompt "
"template; running without it may produce inaccurate relevance "
"scores. Consider specifying `--chat-template %s`.",
self.model_config.model,
suggested_template,
)
#######################################
# online APIs