| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Awaitable |
| 5 | |
| 6 | from fastapi.responses import JSONResponse |
| 7 | from starlette.types import ASGIApp, Receive, Scope, Send |
| 8 | |
| 9 | # Global variable to track scaling state |
| 10 | _scaling_elastic_ep = False |
| 11 | |
| 12 | |
| 13 | def get_scaling_elastic_ep(): |
| 14 | return _scaling_elastic_ep |
| 15 | |
| 16 | |
| 17 | def set_scaling_elastic_ep(value): |
| 18 | global _scaling_elastic_ep |
| 19 | _scaling_elastic_ep = value |
| 20 | |
| 21 | |
| 22 | class ScalingMiddleware: |
| 23 | """Middleware that checks if the model is currently scaling and |
| 24 | returns a 503 Service Unavailable response if it is. |
| 25 | |
| 26 | This middleware applies to all HTTP requests and prevents |
| 27 | processing when the model is in a scaling state. |
| 28 | """ |
| 29 | |
| 30 | def __init__(self, app: ASGIApp) -> None: |
| 31 | self.app = app |
| 32 | |
| 33 | def __call__(self, scope: Scope, receive: Receive, send: Send) -> Awaitable[None]: |
| 34 | if scope["type"] != "http": |
| 35 | return self.app(scope, receive, send) |
| 36 | |
| 37 | # Check global scaling state |
| 38 | if get_scaling_elastic_ep(): |
| 39 | # Return 503 Service Unavailable response |
| 40 | response = JSONResponse( |
| 41 | content={ |
| 42 | "error": "The model is currently scaling. Please try again later." |
| 43 | }, |
| 44 | status_code=503, |
| 45 | ) |
| 46 | return response(scope, receive, send) |
| 47 | |
| 48 | return self.app(scope, receive, send) |