| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import logging |
| 4 | |
| 5 | from fastapi import APIRouter, Depends, FastAPI, Request |
| 6 | from fastapi.responses import JSONResponse, Response |
| 7 | |
| 8 | from vllm import envs |
| 9 | from vllm.entrypoints.openai.models.api_router import models |
| 10 | from vllm.entrypoints.openai.models.serving import OpenAIServingModels |
| 11 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 12 | from vllm.entrypoints.serve.lora.protocol import ( |
| 13 | LoadLoRAAdapterRequest, |
| 14 | UnloadLoRAAdapterRequest, |
| 15 | ) |
| 16 | from vllm.entrypoints.serve.utils.api_utils import validate_json_request |
| 17 | from vllm.logger import init_logger |
| 18 | |
| 19 | logger = init_logger(__name__) |
| 20 | router = APIRouter() |
| 21 | |
| 22 | |
| 23 | def attach_router(app: FastAPI): |
| 24 | """Attach the LoRA adapter load/unload endpoints to the API server. |
| 25 | |
| 26 | Does nothing when dynamic LoRA updating is disabled. Handler levels are |
| 27 | snapshotted and restored because importing |
| 28 | model_hosting_container_standards may reconfigure root logging. |
| 29 | """ |
| 30 | if not envs.VLLM_ALLOW_RUNTIME_LORA_UPDATING: |
| 31 | """If LoRA dynamic loading & unloading is not enabled, do nothing.""" |
| 32 | return |
| 33 | logger.warning( |
| 34 | "LoRA dynamic loading & unloading is enabled in the API server. " |
| 35 | "This should ONLY be used for local development!" |
| 36 | ) |
| 37 | |
| 38 | snapshot = [ |
| 39 | (h, h.level) |
| 40 | for lg in (logging.getLogger(), logging.getLogger("vllm")) |
| 41 | for h in lg.handlers |
| 42 | ] |
| 43 | |
| 44 | try: |
| 45 | _attach_router(app) |
| 46 | finally: |
| 47 | for handler, level in snapshot: |
| 48 | handler.setLevel(level) |
| 49 | |
| 50 | |
| 51 | def _attach_router(app: FastAPI): |
| 52 | """Register the LoRA adapter load/unload routes on the app.""" |
| 53 | import model_hosting_container_standards.sagemaker as sagemaker_standards |
| 54 | |
| 55 | @sagemaker_standards.register_load_adapter_handler( |
| 56 | request_shape={ |
| 57 | "lora_name": "body.name", |
| 58 | "lora_path": "body.src", |
| 59 | "load_inplace": "body.load_inplace || `false`", |
| 60 | "is_3d_lora_weight": "body.is_3d_lora_weight || `false`", |
| 61 | }, |
| 62 | ) |
| 63 | @router.post("/v1/load_lora_adapter", dependencies=[Depends(validate_json_request)]) |
| 64 | async def load_lora_adapter(request: LoadLoRAAdapterRequest, raw_request: Request): |
| 65 | """Handle POST /v1/load_lora_adapter: load a LoRA adapter into the |
| 66 | serving engine.""" |
| 67 | handler: OpenAIServingModels = models(raw_request) |
| 68 | response = await handler.load_lora_adapter(request) |
| 69 | if isinstance(response, ErrorResponse): |
| 70 | return JSONResponse( |
| 71 | content=response.model_dump(), status_code=response.error.code |
| 72 | ) |
| 73 | |
| 74 | return Response(status_code=200, content=response) |
| 75 | |
| 76 | @sagemaker_standards.register_unload_adapter_handler( |
| 77 | request_shape={ |
| 78 | "lora_name": "path_params.adapter_name", |
| 79 | } |
| 80 | ) |
| 81 | @router.post( |
| 82 | "/v1/unload_lora_adapter", dependencies=[Depends(validate_json_request)] |
| 83 | ) |
| 84 | async def unload_lora_adapter( |
| 85 | request: UnloadLoRAAdapterRequest, raw_request: Request |
| 86 | ): |
| 87 | """Handle POST /v1/unload_lora_adapter: unload a LoRA adapter from |
| 88 | the serving engine.""" |
| 89 | handler: OpenAIServingModels = models(raw_request) |
| 90 | response = await handler.unload_lora_adapter(request) |
| 91 | if isinstance(response, ErrorResponse): |
| 92 | return JSONResponse( |
| 93 | content=response.model_dump(), status_code=response.error.code |
| 94 | ) |
| 95 | |
| 96 | return Response(status_code=200, content=response) |
| 97 | |
| 98 | # register the router |
| 99 | app.include_router(router) |