| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from fastapi import APIRouter, Request |
| 5 | from fastapi.responses import JSONResponse |
| 6 | |
| 7 | from vllm.engine.protocol import EngineClient |
| 8 | from vllm.entrypoints.serve.tokenize.serving import ServingTokenization |
| 9 | from vllm.logger import init_logger |
| 10 | from vllm.version import __version__ as VLLM_VERSION |
| 11 | |
| 12 | router = APIRouter() |
| 13 | |
| 14 | logger = init_logger(__name__) |
| 15 | |
| 16 | |
| 17 | def base(request: Request) -> ServingTokenization: |
| 18 | # Reuse the existing instance |
| 19 | return tokenization(request) |
| 20 | |
| 21 | |
| 22 | def tokenization(request: Request) -> ServingTokenization: |
| 23 | return request.app.state.serving_tokenization |
| 24 | |
| 25 | |
| 26 | def engine_client(request: Request) -> EngineClient: |
| 27 | return request.app.state.engine_client |
| 28 | |
| 29 | |
| 30 | @router.get("/load") |
| 31 | async def get_server_load_metrics(request: Request): |
| 32 | # This endpoint returns the current server load metrics. |
| 33 | # It tracks requests utilizing the GPU from the following routes: |
| 34 | # - /v1/responses |
| 35 | # - /v1/responses/{response_id} |
| 36 | # - /v1/responses/{response_id}/cancel |
| 37 | # - /v1/messages |
| 38 | # - /v1/chat/completions |
| 39 | # - /v1/completions |
| 40 | # - /v1/audio/transcriptions |
| 41 | # - /v1/audio/translations |
| 42 | # - /v1/embeddings |
| 43 | # - /pooling |
| 44 | # - /classify |
| 45 | # - /score |
| 46 | # - /v1/score |
| 47 | # - /rerank |
| 48 | # - /v1/rerank |
| 49 | # - /v2/rerank |
| 50 | return JSONResponse(content={"server_load": request.app.state.server_load_metrics}) |
| 51 | |
| 52 | |
| 53 | @router.get("/version") |
| 54 | async def show_version(): |
| 55 | ver = {"version": VLLM_VERSION} |
| 56 | return JSONResponse(content=ver) |
| 57 | |
| 58 | |
| 59 | @router.get("/kv_event_sources") |
| 60 | async def kv_event_sources(raw_request: Request): |
| 61 | """KV-event publisher config per DP rank; empty on render-only servers.""" |
| 62 | client: EngineClient | None = raw_request.app.state.engine_client |
| 63 | return client.get_kv_event_sources() if client is not None else {} |