[Frontend][4/n] Make pooling entrypoints request schema consensus | ScoreRequest (#33060)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
wang.yuqi
2026-02-04 01:48:40 +00:00
committed by GitHub
parent 52ee21021a
commit 1b8fe6f7c4
8 changed files with 432 additions and 205 deletions
+69 -20
View File
@@ -1,5 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from collections.abc import Iterable
from typing import Any, TypeAlias, cast
from torch.nn import CosineSimilarity
@@ -10,12 +11,13 @@ from vllm.entrypoints.chat_utils import (
BaseMultiModalItemTracker,
ChatCompletionContentPartImageEmbedsParam,
ChatCompletionContentPartImageParam,
ChatCompletionContentPartParam,
ChatCompletionContentPartTextParam,
ChatCompletionContentPartVideoParam,
ChatTemplateResolutionError,
ConversationMessage,
MultiModalItemTracker,
_ContentPart,
_parse_chat_message_content_part,
_parse_chat_message_content_parts,
)
from vllm.inputs import TokensPrompt
from vllm.model_executor.models.interfaces import supports_score_template
@@ -46,6 +48,13 @@ class ScoreMultiModalParam(TypedDict, total=False):
"""The multimodal contents"""
# Raw input data with content key in ScoreMultiModalParam.
ScoreInput = str | ScoreMultiModalParam
ScoreInputs = ScoreInput | list[ScoreInput]
# Score data without content key.
ScoreData = str | list[ScoreContentPartParam]
def _cosine_similarity(
tokenizer: TokenizerLike,
embed_1: list[PoolingRequestOutput],
@@ -77,8 +86,8 @@ def _cosine_similarity(
def _validate_score_input_lens(
data_1: list[str] | list[ScoreContentPartParam],
data_2: list[str] | list[ScoreContentPartParam],
data_1: list[ScoreData],
data_2: list[ScoreData],
):
len_1 = len(data_1)
len_2 = len(data_2)
@@ -91,19 +100,56 @@ def _validate_score_input_lens(
raise ValueError("At least one text_pair element must be given")
def _validate_mm_score_input(
data: list[ScoreInput],
is_multimodal_model: bool,
architecture: str,
) -> list[ScoreData]:
out: list[ScoreData] = []
for d in data:
if isinstance(d, str):
out.append(d)
else:
if not is_multimodal_model:
raise ValueError(f"MultiModalParam is not supported for {architecture}")
content = cast(list[ScoreContentPartParam], d.get("content", []))
out.append(content)
return out
def validate_score_input(
data_1: ScoreInputs,
data_2: ScoreInputs,
is_multimodal_model: bool,
architecture: str,
) -> tuple[list[ScoreData], list[ScoreData]]:
if not isinstance(data_1, list):
data_1 = [data_1]
if not isinstance(data_2, list):
data_2 = [data_2]
score_input_1 = _validate_mm_score_input(data_1, is_multimodal_model, architecture)
score_input_2 = _validate_mm_score_input(data_2, is_multimodal_model, architecture)
_validate_score_input_lens(score_input_1, score_input_2)
return score_input_1, score_input_2
def parse_score_data(
data_1: str | ScoreContentPartParam,
data_2: str | ScoreContentPartParam,
data_1: ScoreData,
data_2: ScoreData,
model_config: ModelConfig,
) -> tuple[str, str, MultiModalDataDict | None, MultiModalUUIDDict | None]:
mm_tracker = MultiModalItemTracker(model_config)
content_1 = _parse_score_content(data_1, mm_tracker)
content_2 = _parse_score_content(data_2, mm_tracker)
content_1 = _parse_score_content("query", data_1, mm_tracker)
content_2 = _parse_score_content("document", data_2, mm_tracker)
def ensure_str(content: _ContentPart | None) -> str:
if content is not None and isinstance(content, str):
return cast(str, content)
def ensure_str(content: list[ConversationMessage]) -> str:
assert len(content) == 1
prompt = content[0]["content"]
if prompt is not None and isinstance(prompt, str):
return cast(str, prompt)
else:
raise ValueError(f"Only string content is supported, but got {content}.")
@@ -115,19 +161,22 @@ def parse_score_data(
def _parse_score_content(
data: str | ScoreContentPartParam,
role: str,
data: ScoreData,
mm_tracker: BaseMultiModalItemTracker,
) -> _ContentPart | None:
) -> list[ConversationMessage]:
parts: Iterable[ChatCompletionContentPartParam]
if isinstance(data, str):
part = ChatCompletionContentPartTextParam(type="text", text=data)
parts = [ChatCompletionContentPartTextParam(type="text", text=data)]
else:
part = data
parts = cast(Iterable[ChatCompletionContentPartParam], data)
mm_parser = mm_tracker.create_parser()
parse_res = _parse_chat_message_content_part(
part,
mm_parser,
parse_res = _parse_chat_message_content_parts(
role=role,
parts=parts,
mm_tracker=mm_tracker,
wrap_dicts=False,
interleave_strings=False,
)
@@ -184,8 +233,8 @@ def get_score_prompt(
model_config: ModelConfig,
tokenizer: TokenizerLike,
tokenization_kwargs: dict[str, Any],
data_1: str | ScoreContentPartParam,
data_2: str | ScoreContentPartParam,
data_1: ScoreData,
data_2: ScoreData,
score_template: str | None = None,
) -> tuple[str, TokensPrompt]:
prompt_1, prompt_2, mm_data, mm_uuids = parse_score_data(