| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | |
| 5 | import asyncio |
| 6 | import json |
| 7 | from http import HTTPStatus |
| 8 | |
| 9 | from fastapi import APIRouter, Depends, FastAPI, HTTPException, Request, Response |
| 10 | from fastapi.responses import JSONResponse, StreamingResponse |
| 11 | |
| 12 | from vllm.engine.protocol import EngineClient |
| 13 | from vllm.entrypoints.serve.tokenize.serving import ServingTokenization |
| 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 | from ...serve.engine.protocol import ErrorResponse |
| 22 | from .protocol import ( |
| 23 | GenerateRequest, |
| 24 | GenerateResponse, |
| 25 | ) |
| 26 | from .serving import ServingTokens |
| 27 | |
| 28 | logger = init_logger(__name__) |
| 29 | |
| 30 | |
| 31 | def tokenization(request: Request) -> ServingTokenization: |
| 32 | return request.app.state.serving_tokenization |
| 33 | |
| 34 | |
| 35 | def generate_tokens(request: Request) -> ServingTokens | None: |
| 36 | return request.app.state.serving_tokens |
| 37 | |
| 38 | |
| 39 | def engine_client(request: Request) -> EngineClient: |
| 40 | return request.app.state.engine_client |
| 41 | |
| 42 | |
| 43 | router = APIRouter() |
| 44 | |
| 45 | |
| 46 | @router.post( |
| 47 | "/inference/v1/generate", |
| 48 | dependencies=[Depends(validate_json_request)], |
| 49 | responses={ |
| 50 | HTTPStatus.OK.value: {"content": {"text/event-stream": {}}}, |
| 51 | HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse}, |
| 52 | HTTPStatus.NOT_FOUND.value: {"model": ErrorResponse}, |
| 53 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse}, |
| 54 | }, |
| 55 | ) |
| 56 | @with_cancellation |
| 57 | @load_aware_call |
| 58 | async def generate(request: GenerateRequest, raw_request: Request): |
| 59 | handler = generate_tokens(raw_request) |
| 60 | if handler is None: |
| 61 | raise NotImplementedError("The model does not support generate tokens API") |
| 62 | |
| 63 | generator = await handler.serve_tokens(request, raw_request) |
| 64 | |
| 65 | if isinstance(generator, ErrorResponse): |
| 66 | return JSONResponse( |
| 67 | content=generator.model_dump(), status_code=generator.error.code |
| 68 | ) |
| 69 | |
| 70 | elif isinstance(generator, GenerateResponse): |
| 71 | return JSONResponse(content=generator.model_dump()) |
| 72 | |
| 73 | return StreamingResponse(content=generator, media_type="text/event-stream") |
| 74 | |
| 75 | |
| 76 | def attach_router(app: FastAPI): |
| 77 | if getattr(app.state.args, "tokens_only", False): |
| 78 | |
| 79 | @router.post("/abort_requests") |
| 80 | async def abort_requests(raw_request: Request): |
| 81 | """Abort one or more requests. To be used in a |
| 82 | Disaggregated Everything setup. |
| 83 | """ |
| 84 | try: |
| 85 | body = await raw_request.json() |
| 86 | except json.JSONDecodeError as e: |
| 87 | raise HTTPException( |
| 88 | status_code=HTTPStatus.BAD_REQUEST.value, |
| 89 | detail=f"JSON decode error: {e}", |
| 90 | ) from e |
| 91 | request_ids = body.get("request_ids") |
| 92 | if request_ids is None: |
| 93 | raise HTTPException( |
| 94 | status_code=HTTPStatus.BAD_REQUEST.value, |
| 95 | detail="Missing 'request_ids' in request body", |
| 96 | ) |
| 97 | # Abort requests in background |
| 98 | asyncio.create_task(engine_client(raw_request).abort(request_ids)) |
| 99 | return Response(status_code=200) |
| 100 | |
| 101 | app.include_router(router) |