| 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, FastAPI, Request |
| 6 | from fastapi.responses import JSONResponse |
| 7 | |
| 8 | from vllm.entrypoints.generate.generative_scoring.serving import ( |
| 9 | GenerativeScoringRequest, |
| 10 | GenerativeScoringResponse, |
| 11 | ServingGenerativeScoring, |
| 12 | ) |
| 13 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 14 | from vllm.entrypoints.serve.utils.api_utils import ( |
| 15 | load_aware_call, |
| 16 | validate_json_request, |
| 17 | with_cancellation, |
| 18 | ) |
| 19 | from vllm.logger import init_logger |
| 20 | |
| 21 | router = APIRouter() |
| 22 | |
| 23 | logger = init_logger(__name__) |
| 24 | |
| 25 | |
| 26 | def generative_scoring(request: Request) -> ServingGenerativeScoring | None: |
| 27 | return request.app.state.serving_generative_scoring |
| 28 | |
| 29 | |
| 30 | @router.post( |
| 31 | "/generative_scoring", |
| 32 | dependencies=[Depends(validate_json_request)], |
| 33 | responses={ |
| 34 | HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse}, |
| 35 | HTTPStatus.NOT_FOUND.value: {"model": ErrorResponse}, |
| 36 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse}, |
| 37 | }, |
| 38 | ) |
| 39 | @with_cancellation |
| 40 | @load_aware_call |
| 41 | async def create_generative_scoring( |
| 42 | request: GenerativeScoringRequest, raw_request: Request |
| 43 | ): |
| 44 | handler = generative_scoring(raw_request) |
| 45 | if handler is None: |
| 46 | raise NotImplementedError( |
| 47 | "The model does not support the Generative Scoring API" |
| 48 | ) |
| 49 | |
| 50 | result = await handler.create_generative_scoring(request, raw_request) |
| 51 | |
| 52 | if isinstance(result, ErrorResponse): |
| 53 | return JSONResponse(content=result.model_dump(), status_code=result.error.code) |
| 54 | elif isinstance(result, GenerativeScoringResponse): |
| 55 | return JSONResponse(content=result.model_dump()) |
| 56 | |
| 57 | raise ValueError(f"Unexpected response type: {type(result)}") |
| 58 | |
| 59 | |
| 60 | def register_generative_scoring_api_router(app: FastAPI): |
| 61 | app.include_router(router) |