| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | |
| 5 | from collections.abc import AsyncGenerator |
| 6 | from http import HTTPStatus |
| 7 | |
| 8 | from fastapi import APIRouter, Depends, FastAPI, Request |
| 9 | from fastapi.responses import JSONResponse, StreamingResponse |
| 10 | |
| 11 | from vllm.entrypoints.openai.responses.protocol import ( |
| 12 | ResponsesRequest, |
| 13 | ResponsesResponse, |
| 14 | StreamingResponsesResponse, |
| 15 | ) |
| 16 | from vllm.entrypoints.openai.responses.serving import OpenAIServingResponses |
| 17 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 18 | from vllm.entrypoints.serve.utils.api_utils import ( |
| 19 | load_aware_call, |
| 20 | validate_json_request, |
| 21 | with_cancellation, |
| 22 | ) |
| 23 | from vllm.logger import init_logger |
| 24 | |
| 25 | logger = init_logger(__name__) |
| 26 | |
| 27 | router = APIRouter() |
| 28 | |
| 29 | |
| 30 | def responses(request: Request) -> OpenAIServingResponses | None: |
| 31 | return request.app.state.openai_serving_responses |
| 32 | |
| 33 | |
| 34 | async def _convert_stream_to_sse_events( |
| 35 | generator: AsyncGenerator[StreamingResponsesResponse, None], |
| 36 | ) -> AsyncGenerator[str, None]: |
| 37 | """Convert the generator to a stream of events in SSE format.""" |
| 38 | async for event in generator: |
| 39 | event_type = getattr(event, "type", "unknown") |
| 40 | # https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#event_stream_format |
| 41 | event_data = ( |
| 42 | f"event: {event_type}\ndata: " |
| 43 | f"{event.model_dump_json(indent=None, by_alias=True)}\n\n" |
| 44 | ) |
| 45 | yield event_data |
| 46 | |
| 47 | |
| 48 | @router.post( |
| 49 | "/v1/responses", |
| 50 | dependencies=[Depends(validate_json_request)], |
| 51 | responses={ |
| 52 | HTTPStatus.OK.value: {"content": {"text/event-stream": {}}}, |
| 53 | HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse}, |
| 54 | HTTPStatus.NOT_FOUND.value: {"model": ErrorResponse}, |
| 55 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse}, |
| 56 | }, |
| 57 | ) |
| 58 | @with_cancellation |
| 59 | @load_aware_call |
| 60 | async def create_responses(request: ResponsesRequest, raw_request: Request): |
| 61 | handler = responses(raw_request) |
| 62 | if handler is None: |
| 63 | raise NotImplementedError("The model does not support Responses API") |
| 64 | |
| 65 | generator = await handler.create_responses(request, raw_request) |
| 66 | |
| 67 | if isinstance(generator, ErrorResponse): |
| 68 | return JSONResponse( |
| 69 | content=generator.model_dump(mode="json", by_alias=True), |
| 70 | status_code=generator.error.code, |
| 71 | ) |
| 72 | elif isinstance(generator, ResponsesResponse): |
| 73 | return JSONResponse(content=generator.model_dump(mode="json", by_alias=True)) |
| 74 | |
| 75 | return StreamingResponse( |
| 76 | content=_convert_stream_to_sse_events(generator), media_type="text/event-stream" |
| 77 | ) |
| 78 | |
| 79 | |
| 80 | @router.get("/v1/responses/{response_id}") |
| 81 | @load_aware_call |
| 82 | async def retrieve_responses( |
| 83 | response_id: str, |
| 84 | raw_request: Request, |
| 85 | starting_after: int | None = None, |
| 86 | stream: bool | None = False, |
| 87 | ): |
| 88 | handler = responses(raw_request) |
| 89 | if handler is None: |
| 90 | raise NotImplementedError("The model does not support Responses API") |
| 91 | |
| 92 | response = await handler.retrieve_responses( |
| 93 | response_id, |
| 94 | starting_after=starting_after, |
| 95 | stream=stream, |
| 96 | ) |
| 97 | |
| 98 | if isinstance(response, ErrorResponse): |
| 99 | return JSONResponse( |
| 100 | content=response.model_dump(mode="json", by_alias=True), |
| 101 | status_code=response.error.code, |
| 102 | ) |
| 103 | elif isinstance(response, ResponsesResponse): |
| 104 | return JSONResponse(content=response.model_dump(mode="json", by_alias=True)) |
| 105 | return StreamingResponse( |
| 106 | content=_convert_stream_to_sse_events(response), media_type="text/event-stream" |
| 107 | ) |
| 108 | |
| 109 | |
| 110 | @router.post("/v1/responses/{response_id}/cancel") |
| 111 | @load_aware_call |
| 112 | async def cancel_responses(response_id: str, raw_request: Request): |
| 113 | handler = responses(raw_request) |
| 114 | if handler is None: |
| 115 | raise NotImplementedError("The model does not support Responses API") |
| 116 | |
| 117 | response = await handler.cancel_responses(response_id) |
| 118 | |
| 119 | if isinstance(response, ErrorResponse): |
| 120 | return JSONResponse( |
| 121 | content=response.model_dump(mode="json", by_alias=True), |
| 122 | status_code=response.error.code, |
| 123 | ) |
| 124 | return JSONResponse(content=response.model_dump(mode="json", by_alias=True)) |
| 125 | |
| 126 | |
| 127 | def attach_router(app: FastAPI): |
| 128 | app.include_router(router) |