[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:
@@ -74,3 +74,150 @@ def test_rejects_conflicting_pooling_task(processor: PoolingIOProcessor):
|
|||||||
)
|
)
|
||||||
assert exc_info.value.parameter is None
|
assert exc_info.value.parameter is None
|
||||||
assert exc_info.value.value is None
|
assert exc_info.value.value is None
|
||||||
|
|
||||||
|
|
||||||
|
def test_qwen3_reranker_warns_without_chat_template(monkeypatch):
|
||||||
|
from unittest.mock import MagicMock
|
||||||
|
|
||||||
|
from vllm.entrypoints.pooling.scoring.io_processor import CrossEncoderIOProcessor
|
||||||
|
|
||||||
|
monkeypatch.setattr(
|
||||||
|
"vllm.model_executor.model_loader.get_model_cls", lambda *_: MagicMock()
|
||||||
|
)
|
||||||
|
monkeypatch.setattr(
|
||||||
|
"vllm.model_executor.models.interfaces.supports_score_template",
|
||||||
|
lambda *_: False,
|
||||||
|
)
|
||||||
|
monkeypatch.setattr(
|
||||||
|
"vllm.entrypoints.pooling.scoring.io_processor.is_mistral_tokenizer",
|
||||||
|
lambda *_: False,
|
||||||
|
)
|
||||||
|
|
||||||
|
mock_logger = MagicMock()
|
||||||
|
monkeypatch.setattr(
|
||||||
|
"vllm.entrypoints.pooling.scoring.io_processor.logger",
|
||||||
|
mock_logger,
|
||||||
|
)
|
||||||
|
|
||||||
|
model_config = MagicMock()
|
||||||
|
model_config.architecture = "Qwen3ForSequenceClassification"
|
||||||
|
model_config.model = "Qwen/Qwen3-Reranker-0.6B"
|
||||||
|
model_config.hf_config.is_original_qwen3_reranker = True
|
||||||
|
model_config.use_sep_token = False
|
||||||
|
model_config.is_multimodal_model = False
|
||||||
|
model_config.max_model_len = 8192
|
||||||
|
|
||||||
|
vllm_config = MagicMock(model_config=model_config)
|
||||||
|
tokenizer = MagicMock()
|
||||||
|
tokenizer.pad_token_id = 0
|
||||||
|
renderer = MagicMock()
|
||||||
|
renderer.get_tokenizer.return_value = tokenizer
|
||||||
|
chat_template_config = MagicMock(chat_template=None)
|
||||||
|
|
||||||
|
CrossEncoderIOProcessor(
|
||||||
|
vllm_config=vllm_config,
|
||||||
|
renderer=renderer,
|
||||||
|
chat_template_config=chat_template_config,
|
||||||
|
)
|
||||||
|
|
||||||
|
mock_logger.warning.assert_called_once()
|
||||||
|
warning_call = mock_logger.warning.call_args
|
||||||
|
assert "Qwen/Qwen3-Reranker-0.6B" in warning_call[0][1]
|
||||||
|
assert "qwen3_reranker.jinja" in warning_call[0][2]
|
||||||
|
|
||||||
|
|
||||||
|
def test_qwen3_vl_reranker_warns_with_vl_template(monkeypatch):
|
||||||
|
from unittest.mock import MagicMock
|
||||||
|
|
||||||
|
from vllm.entrypoints.pooling.scoring.io_processor import CrossEncoderIOProcessor
|
||||||
|
|
||||||
|
monkeypatch.setattr(
|
||||||
|
"vllm.model_executor.model_loader.get_model_cls", lambda *_: MagicMock()
|
||||||
|
)
|
||||||
|
monkeypatch.setattr(
|
||||||
|
"vllm.model_executor.models.interfaces.supports_score_template",
|
||||||
|
lambda *_: False,
|
||||||
|
)
|
||||||
|
monkeypatch.setattr(
|
||||||
|
"vllm.entrypoints.pooling.scoring.io_processor.is_mistral_tokenizer",
|
||||||
|
lambda *_: False,
|
||||||
|
)
|
||||||
|
|
||||||
|
mock_logger = MagicMock()
|
||||||
|
monkeypatch.setattr(
|
||||||
|
"vllm.entrypoints.pooling.scoring.io_processor.logger",
|
||||||
|
mock_logger,
|
||||||
|
)
|
||||||
|
|
||||||
|
model_config = MagicMock()
|
||||||
|
model_config.architecture = "Qwen3VLForSequenceClassification"
|
||||||
|
model_config.model = "Qwen/Qwen3-VL-Reranker-2B"
|
||||||
|
model_config.hf_config.is_original_qwen3_reranker = True
|
||||||
|
model_config.use_sep_token = False
|
||||||
|
model_config.is_multimodal_model = True
|
||||||
|
model_config.max_model_len = 8192
|
||||||
|
|
||||||
|
vllm_config = MagicMock(model_config=model_config)
|
||||||
|
tokenizer = MagicMock()
|
||||||
|
tokenizer.pad_token_id = 0
|
||||||
|
renderer = MagicMock()
|
||||||
|
renderer.get_tokenizer.return_value = tokenizer
|
||||||
|
chat_template_config = MagicMock(chat_template=None)
|
||||||
|
|
||||||
|
CrossEncoderIOProcessor(
|
||||||
|
vllm_config=vllm_config,
|
||||||
|
renderer=renderer,
|
||||||
|
chat_template_config=chat_template_config,
|
||||||
|
)
|
||||||
|
|
||||||
|
mock_logger.warning.assert_called_once()
|
||||||
|
warning_call = mock_logger.warning.call_args
|
||||||
|
assert "Qwen/Qwen3-VL-Reranker-2B" in warning_call[0][1]
|
||||||
|
assert "qwen3_vl_reranker.jinja" in warning_call[0][2]
|
||||||
|
|
||||||
|
|
||||||
|
def test_qwen3_reranker_no_warning_when_template_provided(monkeypatch):
|
||||||
|
from unittest.mock import MagicMock
|
||||||
|
|
||||||
|
from vllm.entrypoints.pooling.scoring.io_processor import CrossEncoderIOProcessor
|
||||||
|
|
||||||
|
monkeypatch.setattr(
|
||||||
|
"vllm.model_executor.model_loader.get_model_cls", lambda *_: MagicMock()
|
||||||
|
)
|
||||||
|
monkeypatch.setattr(
|
||||||
|
"vllm.model_executor.models.interfaces.supports_score_template",
|
||||||
|
lambda *_: False,
|
||||||
|
)
|
||||||
|
monkeypatch.setattr(
|
||||||
|
"vllm.entrypoints.pooling.scoring.io_processor.is_mistral_tokenizer",
|
||||||
|
lambda *_: False,
|
||||||
|
)
|
||||||
|
|
||||||
|
mock_logger = MagicMock()
|
||||||
|
monkeypatch.setattr(
|
||||||
|
"vllm.entrypoints.pooling.scoring.io_processor.logger",
|
||||||
|
mock_logger,
|
||||||
|
)
|
||||||
|
|
||||||
|
model_config = MagicMock()
|
||||||
|
model_config.architecture = "Qwen3ForSequenceClassification"
|
||||||
|
model_config.model = "Qwen/Qwen3-Reranker-0.6B"
|
||||||
|
model_config.hf_config.is_original_qwen3_reranker = True
|
||||||
|
model_config.use_sep_token = False
|
||||||
|
model_config.is_multimodal_model = False
|
||||||
|
model_config.max_model_len = 8192
|
||||||
|
|
||||||
|
vllm_config = MagicMock(model_config=model_config)
|
||||||
|
tokenizer = MagicMock()
|
||||||
|
tokenizer.pad_token_id = 0
|
||||||
|
renderer = MagicMock()
|
||||||
|
renderer.get_tokenizer.return_value = tokenizer
|
||||||
|
chat_template_config = MagicMock(chat_template="custom template")
|
||||||
|
|
||||||
|
CrossEncoderIOProcessor(
|
||||||
|
vllm_config=vllm_config,
|
||||||
|
renderer=renderer,
|
||||||
|
chat_template_config=chat_template_config,
|
||||||
|
)
|
||||||
|
|
||||||
|
mock_logger.warning.assert_not_called()
|
||||||
|
|||||||
@@ -7,6 +7,7 @@ from typing import Any, TypeAlias, cast
|
|||||||
import torch.nn.functional as F
|
import torch.nn.functional as F
|
||||||
|
|
||||||
from vllm import PoolingParams, PoolingRequestOutput, TokensPrompt
|
from vllm import PoolingParams, PoolingRequestOutput, TokensPrompt
|
||||||
|
from vllm.logger import init_logger
|
||||||
from vllm.renderers import TokenizeParams
|
from vllm.renderers import TokenizeParams
|
||||||
from vllm.renderers.hf import safe_apply_chat_template
|
from vllm.renderers.hf import safe_apply_chat_template
|
||||||
from vllm.renderers.inputs.preprocess import (
|
from vllm.renderers.inputs.preprocess import (
|
||||||
@@ -47,6 +48,8 @@ from .utils import (
|
|||||||
validate_score_input,
|
validate_score_input,
|
||||||
)
|
)
|
||||||
|
|
||||||
|
logger = init_logger(__name__)
|
||||||
|
|
||||||
ScoringServeContext: TypeAlias = PoolingServeContext[ScoringRequest]
|
ScoringServeContext: TypeAlias = PoolingServeContext[ScoringRequest]
|
||||||
|
|
||||||
|
|
||||||
@@ -425,6 +428,24 @@ class CrossEncoderIOProcessor(ScoringIOProcessor):
|
|||||||
self.model = model if self.supports_score_template else None
|
self.model = model if self.supports_score_template else None
|
||||||
self.use_sep_token = self.model_config.use_sep_token
|
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
|
# online APIs
|
||||||
|
|
||||||
|
|||||||
Reference in New Issue
Block a user