| 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 | from fastapi.responses import JSONResponse |
| 7 | |
| 8 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 9 | from vllm.entrypoints.serve.utils.api_utils import validate_json_request |
| 10 | from vllm.logger import init_logger |
| 11 | |
| 12 | from ..token_in_token_out.protocol import ( |
| 13 | DerenderChatRequestUnion, |
| 14 | DerenderChatStreamRequest, |
| 15 | DerenderChatStreamResponse, |
| 16 | DerenderCompletionRequestUnion, |
| 17 | DerenderCompletionStreamRequest, |
| 18 | DerenderCompletionStreamResponse, |
| 19 | ) |
| 20 | from .serving import ServingDerender |
| 21 | |
| 22 | logger = init_logger(__name__) |
| 23 | |
| 24 | router = APIRouter() |
| 25 | |
| 26 | |
| 27 | def derender(request: Request) -> ServingDerender | None: |
| 28 | return getattr(request.app.state, "serving_derender", None) |
| 29 | |
| 30 | |
| 31 | @router.post( |
| 32 | "/v1/chat/completions/derender", |
| 33 | dependencies=[Depends(validate_json_request)], |
| 34 | responses={ |
| 35 | HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse}, |
| 36 | HTTPStatus.NOT_FOUND.value: {"model": ErrorResponse}, |
| 37 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse}, |
| 38 | }, |
| 39 | ) |
| 40 | async def derender_chat_completion( |
| 41 | request: DerenderChatRequestUnion, |
| 42 | raw_request: Request, |
| 43 | ): |
| 44 | """Derender a generate response into a ChatCompletionResponse. |
| 45 | |
| 46 | Accepts both non-streaming (``stream=false``, default) and streaming |
| 47 | (``stream=true``) request bodies on the same path; FastAPI validates and |
| 48 | routes on the ``stream`` discriminator. |
| 49 | |
| 50 | Non-streaming: body is ``DerenderChatRequest`` (``generate_response`` with |
| 51 | the complete token list). Returns a ``ChatCompletionResponse``. |
| 52 | |
| 53 | Streaming: body is ``DerenderChatStreamRequest`` (one ``generate_chunk`` |
| 54 | delta + optional ``stream_state``). Returns a ``DerenderChatStreamResponse`` |
| 55 | (``chunk`` + ``stream_state``). The client carries ``stream_state`` between |
| 56 | successive calls, one per SSE chunk from ``/inference/v1/generate``. |
| 57 | """ |
| 58 | handler = derender(raw_request) |
| 59 | if handler is None: |
| 60 | raise NotImplementedError( |
| 61 | "The model does not support Chat Completions Derender API" |
| 62 | ) |
| 63 | |
| 64 | if isinstance(request, DerenderChatStreamRequest): |
| 65 | stream_result = await handler.derender_chat_stream_response(request) |
| 66 | if isinstance(stream_result, ErrorResponse): |
| 67 | return JSONResponse( |
| 68 | content=stream_result.model_dump(), |
| 69 | status_code=stream_result.error.code, |
| 70 | ) |
| 71 | chunk, stream_state = stream_result |
| 72 | response = DerenderChatStreamResponse(chunk=chunk, stream_state=stream_state) |
| 73 | return JSONResponse(content=response.model_dump()) |
| 74 | |
| 75 | result = await handler.derender_chat_response(request) |
| 76 | if isinstance(result, ErrorResponse): |
| 77 | return JSONResponse(content=result.model_dump(), status_code=result.error.code) |
| 78 | return JSONResponse(content=result.model_dump()) |
| 79 | |
| 80 | |
| 81 | @router.post( |
| 82 | "/v1/completions/derender", |
| 83 | dependencies=[Depends(validate_json_request)], |
| 84 | responses={ |
| 85 | HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse}, |
| 86 | HTTPStatus.NOT_FOUND.value: {"model": ErrorResponse}, |
| 87 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse}, |
| 88 | }, |
| 89 | ) |
| 90 | async def derender_completion( |
| 91 | request: DerenderCompletionRequestUnion, |
| 92 | raw_request: Request, |
| 93 | ): |
| 94 | """Derender a generate response into a CompletionResponse. |
| 95 | |
| 96 | Accepts both non-streaming (``stream=false``, default) and streaming |
| 97 | (``stream=true``) request bodies on the same path. |
| 98 | |
| 99 | Non-streaming: body is ``DerenderCompletionRequest``. Returns a |
| 100 | ``CompletionResponse``. |
| 101 | |
| 102 | Streaming: body is ``DerenderCompletionStreamRequest`` (one |
| 103 | ``generate_chunk`` + optional ``stream_state``). Returns a |
| 104 | ``DerenderCompletionStreamResponse`` (``chunk`` + ``stream_state``). |
| 105 | """ |
| 106 | handler = derender(raw_request) |
| 107 | if handler is None: |
| 108 | raise NotImplementedError("The model does not support Completions Derender API") |
| 109 | |
| 110 | if isinstance(request, DerenderCompletionStreamRequest): |
| 111 | stream_result = await handler.derender_completion_stream_response(request) |
| 112 | if isinstance(stream_result, ErrorResponse): |
| 113 | return JSONResponse( |
| 114 | content=stream_result.model_dump(), |
| 115 | status_code=stream_result.error.code, |
| 116 | ) |
| 117 | chunk, stream_state = stream_result |
| 118 | response = DerenderCompletionStreamResponse( |
| 119 | chunk=chunk, stream_state=stream_state |
| 120 | ) |
| 121 | return JSONResponse(content=response.model_dump()) |
| 122 | |
| 123 | result = await handler.derender_completion_response(request) |
| 124 | if isinstance(result, ErrorResponse): |
| 125 | return JSONResponse(content=result.model_dump(), status_code=result.error.code) |
| 126 | return JSONResponse(content=result.model_dump()) |