| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from http import HTTPStatus |
| 4 | |
| 5 | from fastapi import APIRouter, Depends, Request |
| 6 | |
| 7 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 8 | from vllm.entrypoints.serve.utils.api_utils import ( |
| 9 | load_aware_call, |
| 10 | validate_json_request, |
| 11 | with_cancellation, |
| 12 | ) |
| 13 | from vllm.logger import init_logger |
| 14 | |
| 15 | from .protocol import RerankRequest, ScoreRequest |
| 16 | from .serving import ServingScores |
| 17 | |
| 18 | router = APIRouter() |
| 19 | |
| 20 | logger = init_logger(__name__) |
| 21 | |
| 22 | |
| 23 | def score(request: Request) -> ServingScores: |
| 24 | handler = getattr(request.app.state, "serving_scores", None) |
| 25 | if handler is None: |
| 26 | raise NotImplementedError("The model does not support Score API") |
| 27 | return handler |
| 28 | |
| 29 | |
| 30 | def rerank(request: Request) -> ServingScores: |
| 31 | handler = getattr(request.app.state, "serving_scores", None) |
| 32 | if handler is None: |
| 33 | raise NotImplementedError("The model does not support Rerank (Score) API") |
| 34 | return handler |
| 35 | |
| 36 | |
| 37 | @router.post( |
| 38 | "/score", |
| 39 | dependencies=[Depends(validate_json_request)], |
| 40 | responses={ |
| 41 | HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse}, |
| 42 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse}, |
| 43 | }, |
| 44 | ) |
| 45 | @with_cancellation |
| 46 | @load_aware_call |
| 47 | async def create_score(request: ScoreRequest, raw_request: Request): |
| 48 | handler = score(raw_request) |
| 49 | return await handler(request, raw_request) |
| 50 | |
| 51 | |
| 52 | @router.post( |
| 53 | "/v1/score", |
| 54 | dependencies=[Depends(validate_json_request)], |
| 55 | responses={ |
| 56 | HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse}, |
| 57 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse}, |
| 58 | }, |
| 59 | ) |
| 60 | async def create_score_v1(request: ScoreRequest, raw_request: Request): |
| 61 | logger.warning( |
| 62 | "To indicate that Score API is not part of standard OpenAI API, we " |
| 63 | "have moved it to `/score`. Please update your client accordingly." |
| 64 | ) |
| 65 | |
| 66 | return await create_score(request, raw_request) |
| 67 | |
| 68 | |
| 69 | @router.post( |
| 70 | "/rerank", |
| 71 | dependencies=[Depends(validate_json_request)], |
| 72 | responses={ |
| 73 | HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse}, |
| 74 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse}, |
| 75 | }, |
| 76 | ) |
| 77 | @with_cancellation |
| 78 | @load_aware_call |
| 79 | async def do_rerank(request: RerankRequest, raw_request: Request): |
| 80 | handler = rerank(raw_request) |
| 81 | return await handler(request, raw_request) |
| 82 | |
| 83 | |
| 84 | @router.post( |
| 85 | "/v1/rerank", |
| 86 | dependencies=[Depends(validate_json_request)], |
| 87 | responses={ |
| 88 | HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse}, |
| 89 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse}, |
| 90 | }, |
| 91 | ) |
| 92 | async def do_rerank_v1(request: RerankRequest, raw_request: Request): |
| 93 | logger.warning_once( |
| 94 | "To indicate that the rerank API is not part of the standard OpenAI" |
| 95 | " API, we have located it at `/rerank`. Please update your client " |
| 96 | "accordingly. (Note: Conforms to JinaAI rerank API)" |
| 97 | ) |
| 98 | |
| 99 | return await do_rerank(request, raw_request) |
| 100 | |
| 101 | |
| 102 | @router.post( |
| 103 | "/v2/rerank", |
| 104 | dependencies=[Depends(validate_json_request)], |
| 105 | responses={ |
| 106 | HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse}, |
| 107 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse}, |
| 108 | }, |
| 109 | ) |
| 110 | async def do_rerank_v2(request: RerankRequest, raw_request: Request): |
| 111 | return await do_rerank(request, raw_request) |