| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from http import HTTPStatus |
| 5 | |
| 6 | from fastapi import APIRouter, Depends, Request |
| 7 | |
| 8 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 9 | from vllm.entrypoints.serve.utils.api_utils import ( |
| 10 | load_aware_call, |
| 11 | validate_json_request, |
| 12 | with_cancellation, |
| 13 | ) |
| 14 | |
| 15 | from .protocol import CohereEmbedRequest, EmbeddingRequest |
| 16 | from .serving import ServingEmbedding |
| 17 | |
| 18 | router = APIRouter() |
| 19 | |
| 20 | |
| 21 | def embedding(request: Request) -> ServingEmbedding: |
| 22 | handler = getattr(request.app.state, "serving_embedding", None) |
| 23 | if handler is None: |
| 24 | raise NotImplementedError("The model does not support Embeddings API") |
| 25 | return handler |
| 26 | |
| 27 | |
| 28 | @router.post( |
| 29 | "/v1/embeddings", |
| 30 | dependencies=[Depends(validate_json_request)], |
| 31 | responses={ |
| 32 | HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse}, |
| 33 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse}, |
| 34 | }, |
| 35 | ) |
| 36 | @with_cancellation |
| 37 | @load_aware_call |
| 38 | async def create_embedding( |
| 39 | request: EmbeddingRequest, |
| 40 | raw_request: Request, |
| 41 | ): |
| 42 | handler = embedding(raw_request) |
| 43 | return await handler(request, raw_request) |
| 44 | |
| 45 | |
| 46 | @router.post( |
| 47 | "/v2/embed", |
| 48 | dependencies=[Depends(validate_json_request)], |
| 49 | responses={ |
| 50 | HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse}, |
| 51 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse}, |
| 52 | }, |
| 53 | ) |
| 54 | @with_cancellation |
| 55 | @load_aware_call |
| 56 | async def create_cohere_embedding( |
| 57 | request: CohereEmbedRequest, |
| 58 | raw_request: Request, |
| 59 | ): |
| 60 | handler = embedding(raw_request) |
| 61 | return await handler(request, raw_request) |