| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from dataclasses import dataclass |
| 4 | from typing import TypeAlias |
| 5 | |
| 6 | from typing_extensions import Required, TypedDict |
| 7 | |
| 8 | from vllm.entrypoints.chat_utils import ( |
| 9 | ChatCompletionContentPartImageEmbedsParam, |
| 10 | ChatCompletionContentPartImageParam, |
| 11 | ChatCompletionContentPartTextParam, |
| 12 | ChatCompletionContentPartVideoParam, |
| 13 | ) |
| 14 | |
| 15 | ScoreContentPartParam: TypeAlias = ( |
| 16 | ChatCompletionContentPartImageParam |
| 17 | | ChatCompletionContentPartImageEmbedsParam |
| 18 | | ChatCompletionContentPartTextParam |
| 19 | | ChatCompletionContentPartVideoParam |
| 20 | ) |
| 21 | |
| 22 | |
| 23 | class ScoreMultiModalParam(TypedDict, total=False): |
| 24 | """A specialized parameter type for scoring multimodal content. |
| 25 | |
| 26 | The reasons why don't reuse `CustomChatCompletionMessageParam` directly: |
| 27 | 1. Score tasks don't need the 'role' field (user/assistant/system) that's required in chat completions |
| 28 | 2. Including chat-specific fields would confuse users about their purpose in scoring |
| 29 | 3. This is a more focused interface that only exposes what's needed for scoring |
| 30 | """ # noqa: E501 |
| 31 | |
| 32 | content: Required[list[ScoreContentPartParam]] |
| 33 | """The multimodal contents""" |
| 34 | |
| 35 | |
| 36 | # Raw input data with content key in ScoreMultiModalParam. |
| 37 | ScoreInput = str | ScoreMultiModalParam |
| 38 | # Score data without content key. |
| 39 | ScoreData = str | list[ScoreContentPartParam] |
| 40 | |
| 41 | |
| 42 | @dataclass |
| 43 | class ScoringData: |
| 44 | data_1: list[ScoreData] |
| 45 | data_2: list[ScoreData] |