[Frontend] Support using chat template as custom score template for reranking models (#30550)

Signed-off-by: Jakub Zakrzewski <jzakrzewski@nvidia.com>
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
Signed-off-by: wang.yuqi <noooop@126.com>
Co-authored-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
Jakub Zakrzewski
2025-12-23 11:19:16 +00:00
committed by GitHub
co-authored by wang.yuqi
parent 27c6c2f98c
commit 23daef548d
19 changed files with 663 additions and 46 deletions
+10 -1
View File
@@ -67,6 +67,15 @@ else:
logger = init_logger(__name__)
class ChatTemplateResolutionError(ValueError):
"""Raised when chat template resolution fails.
This is a subclass of ValueError for backward compatibility with
existing exception handlers.
"""
MODALITY_PLACEHOLDERS_MAP = {
"image": "<##IMAGE##>",
"audio": "<##AUDIO##>",
@@ -1814,7 +1823,7 @@ def apply_hf_chat_template(
)
if hf_chat_template is None:
raise ValueError(
raise ChatTemplateResolutionError(
"As of transformers v4.44, default chat template is no longer "
"allowed, so you must provide a chat template if the tokenizer "
"does not define one."
+11
View File
@@ -1280,6 +1280,7 @@ class LLM:
pooling_params: PoolingParams | None = None,
lora_request: list[LoRARequest] | LoRARequest | None = None,
tokenization_kwargs: dict[str, Any] | None = None,
score_template: str | None = None,
) -> list[ScoringRequestOutput]:
model_config = self.model_config
@@ -1313,6 +1314,7 @@ class LLM:
data_2=d,
tokenizer=tokenizer,
tokenization_kwargs=tokenization_kwargs,
score_template=score_template,
)
if token_type_ids := engine_prompt.pop("token_type_ids", None):
@@ -1347,6 +1349,7 @@ class LLM:
use_tqdm: bool | Callable[..., tqdm] = True,
pooling_params: PoolingParams | None = None,
lora_request: list[LoRARequest] | LoRARequest | None = None,
chat_template: str | None = None,
) -> list[ScoringRequestOutput]:
"""Generate similarity scores for all pairs `<text,text_pair>` or
`<multi-modal data, multi-modal data pair>`.
@@ -1379,6 +1382,8 @@ class LLM:
lora_request: LoRA request to use for generation, if any.
pooling_params: The pooling parameters for pooling. If None, we
use the default pooling parameters.
chat_template: The chat template to use for the scoring. If None, we
use the model's default chat template.
Returns:
A list of `ScoringRequestOutput` objects containing the
generated scores in the same order as the input prompts.
@@ -1406,6 +1411,11 @@ class LLM:
):
raise ValueError("Score API is only enabled for num_labels == 1.")
if not model_config.is_cross_encoder and chat_template is not None:
raise ValueError(
"chat_template is only supported for cross-encoder models."
)
# the tokenizer for models such as
# "cross-encoder/ms-marco-MiniLM-L-6-v2" doesn't support passing
# lists of tokens to the `text` and `text_pair` kwargs
@@ -1475,6 +1485,7 @@ class LLM:
use_tqdm,
pooling_params,
lora_request,
score_template=chat_template,
)
else:
return self._embedding_score(
+1
View File
@@ -1145,6 +1145,7 @@ async def init_app_state(
engine_client,
state.openai_serving_models,
request_logger=request_logger,
score_template=resolved_chat_template,
log_error_stack=args.log_error_stack,
)
if ("embed" in supported_tasks or "score" in supported_tasks)
+1
View File
@@ -495,6 +495,7 @@ async def run_batch(
engine_client,
openai_serving_models,
request_logger=request_logger,
score_template=None,
)
if ("embed" in supported_tasks or enable_serving_reranking)
else None
@@ -52,6 +52,7 @@ class ServingScores(OpenAIServing):
models: OpenAIServingModels,
*,
request_logger: RequestLogger | None,
score_template: str | None = None,
log_error_stack: bool = False,
) -> None:
super().__init__(
@@ -60,6 +61,7 @@ class ServingScores(OpenAIServing):
request_logger=request_logger,
log_error_stack=log_error_stack,
)
self.score_template = score_template
async def _embedding_score(
self,
@@ -169,6 +171,7 @@ class ServingScores(OpenAIServing):
data_2=data_2,
tokenizer=tokenizer,
tokenization_kwargs=tokenization_kwargs,
score_template=self.score_template,
)
self._validate_input(request, engine_prompt["prompt_token_ids"], full_prompt)
if request.mm_processor_kwargs is not None:
+46 -16
View File
@@ -11,9 +11,11 @@ from vllm.entrypoints.chat_utils import (
ChatCompletionContentPartImageEmbedsParam,
ChatCompletionContentPartImageParam,
ChatCompletionContentPartTextParam,
ChatTemplateResolutionError,
MultiModalItemTracker,
_ContentPart,
_parse_chat_message_content_part,
apply_hf_chat_template,
)
from vllm.inputs import TokensPrompt
from vllm.model_executor.models.interfaces import supports_score_template
@@ -139,10 +141,8 @@ def _parse_score_content(
return next(iter(mm_placeholder_storage.values()))[0]
def apply_score_template(
model_config: ModelConfig,
prompt_1: str,
prompt_2: str,
def _apply_model_score_template(
model_config: ModelConfig, prompt_1: str, prompt_2: str
) -> str:
# NOTE(Simon): lazy import to avoid bring in all dependencies (e.g. gguf)
from vllm.model_executor.model_loader import get_model_cls
@@ -181,6 +181,7 @@ def get_score_prompt(
tokenization_kwargs: dict[str, Any],
data_1: str | ScoreContentPartParam,
data_2: str | ScoreContentPartParam,
score_template: str | None = None,
) -> tuple[str, TokensPrompt]:
prompt_1, prompt_2, mm_data = parse_score_data(
data_1,
@@ -190,19 +191,48 @@ def get_score_prompt(
from vllm.model_executor.model_loader import get_model_cls
model = get_model_cls(model_config)
if supports_score_template(model):
full_prompt = apply_score_template(model_config, prompt_1, prompt_2)
prompt_inputs = tokenizer(full_prompt, **tokenization_kwargs)
elif model_config.use_pad_token:
# cross_encoder models defaults to using pad_token.
prompt_inputs = tokenizer(
text=prompt_1, text_pair=prompt_2, **tokenization_kwargs
)
full_prompt = tokenizer.decode(prompt_inputs["input_ids"])
def default_tokenizer_encode():
if supports_score_template(model):
full_prompt = _apply_model_score_template(model_config, prompt_1, prompt_2)
prompt_inputs = tokenizer(full_prompt, **tokenization_kwargs)
else:
if model_config.use_pad_token:
# cross_encoder models defaults to using pad_token.
prompt_inputs = tokenizer(
text=prompt_1, text_pair=prompt_2, **tokenization_kwargs
)
full_prompt = tokenizer.decode(prompt_inputs["input_ids"])
else:
# `llm as reranker` models defaults to not using pad_token.
full_prompt = prompt_1 + prompt_2
prompt_inputs = tokenizer(text=full_prompt, **tokenization_kwargs)
return full_prompt, prompt_inputs
# FIXME: For now, we only apply a template when one is explicitly provided.
# We cannot rely on the tokenizer's chat template because many models
# inherit junk templates from their base LLM, which breaks both the models
# and the tests that use them.
if score_template is None:
full_prompt, prompt_inputs = default_tokenizer_encode()
else:
# `llm as reranker` models defaults to not using pad_token.
full_prompt = prompt_1 + prompt_2
prompt_inputs = tokenizer(text=full_prompt, **tokenization_kwargs)
# FIXME: Try applying a score template from the CLI arg or tokenizer_config.json
# If that fails because there is no such template,
# fall back to the default implementation.
try:
full_prompt = apply_hf_chat_template(
tokenizer,
[
{"role": "query", "content": prompt_1},
{"role": "document", "content": prompt_2},
],
score_template,
tools=None,
model_config=model_config,
)
prompt_inputs = tokenizer(full_prompt, **tokenization_kwargs)
except ChatTemplateResolutionError:
full_prompt, prompt_inputs = default_tokenizer_encode()
engine_prompt = TokensPrompt(prompt_token_ids=prompt_inputs["input_ids"])
+22
View File
@@ -88,6 +88,26 @@ class JinaRobertaModelConfig(VerifyAndUpdateConfig):
}
class LlamaBidirectionalConfig(VerifyAndUpdateConfig):
@staticmethod
def verify_and_update_config(vllm_config: "VllmConfig") -> None:
from vllm.config.pooler import PoolingTypeStr
hf_config = vllm_config.model_config.hf_config
hf_config.is_causal = False
pooling_type_map: dict[str, PoolingTypeStr] = {
"avg": "MEAN",
"cls": "CLS",
"last": "LAST",
}
pooling_type = pooling_type_map.get(hf_config.pooling, None)
if pooling_type is None:
raise ValueError(f"pool_type {hf_config.pooling} not supported")
vllm_config.model_config.pooler_config.pooling_type = pooling_type
class NomicBertModelConfig(VerifyAndUpdateConfig):
@staticmethod
def verify_and_update_config(vllm_config: "VllmConfig") -> None:
@@ -509,6 +529,8 @@ MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = {
"GteNewModel": GteNewModelConfig,
"GteNewForSequenceClassification": GteNewModelConfig,
"Gemma3TextModel": Gemma3TextModelConfig,
"LlamaBidirectionalForSequenceClassification": LlamaBidirectionalConfig,
"LlamaBidirectionalModel": LlamaBidirectionalConfig,
"NomicBertModel": NomicBertModelConfig,
"Qwen2ForProcessRewardModel": Qwen2ForProcessRewardModelConfig,
"Qwen2ForRewardModel": Qwen2ForRewardModelConfig,
+22 -1
View File
@@ -57,7 +57,14 @@ from vllm.model_executor.model_loader.weight_utils import (
)
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsEagle, SupportsEagle3, SupportsLoRA, SupportsPP
from .adapters import as_embedding_model, as_seq_cls_model
from .interfaces import (
SupportsEagle,
SupportsEagle3,
SupportsLoRA,
SupportsPP,
)
from .interfaces_base import attn_type
from .utils import (
AutoWeightsLoader,
PPMissingLayer,
@@ -698,3 +705,17 @@ class LlamaForCausalLM(
name = name.replace(item, mapping[item])
return name, loaded_weight
@attn_type("encoder_only")
class LlamaBidirectionalForSequenceClassification(as_seq_cls_model(LlamaForCausalLM)):
# This class sets the correct attention type and pooling type
# through LlamaBidirectionalConfig.
pass
@attn_type("encoder_only")
class LlamaBidirectionalModel(as_embedding_model(LlamaForCausalLM)):
# This class sets the correct attention type and pooling type
# through LlamaBidirectionalConfig.
pass
+6 -2
View File
@@ -203,6 +203,7 @@ _EMBEDDING_MODELS = {
"GteNewModel": ("bert_with_rope", "GteNewModel"),
"InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
"JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"), # noqa: E501
"LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
"LlamaModel": ("llama", "LlamaForCausalLM"),
**{
# Multiple models share the same architecture, so we include them all
@@ -246,6 +247,11 @@ _CROSS_ENCODER_MODELS = {
"bert_with_rope",
"GteNewForSequenceClassification",
),
"JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
"LlamaBidirectionalForSequenceClassification": (
"llama",
"LlamaBidirectionalForSequenceClassification",
),
"ModernBertForSequenceClassification": (
"modernbert",
"ModernBertForSequenceClassification",
@@ -259,8 +265,6 @@ _CROSS_ENCODER_MODELS = {
"roberta",
"RobertaForSequenceClassification",
),
# [Auto-converted (see adapters.py)]
"JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"), # noqa: E501,
}
_MULTIMODAL_MODELS = {