multi-LoRA as extra models in OpenAI server (#2775)

how to serve the loras (mimicking the [multilora inference example](https://github.com/vllm-project/vllm/blob/main/examples/multilora_inference.py)):
```terminal
$ export LORA_PATH=~/.cache/huggingface/hub/models--yard1--llama-2-7b-sql-lora-test/
$ python -m vllm.entrypoints.api_server \
 --model meta-llama/Llama-2-7b-hf \
 --enable-lora \
 --lora-modules sql-lora=$LORA_PATH sql-lora2=$LORA_PATH
```
the above server will list 3 separate values if the user queries `/models`: one for the base served model, and one each for the specified lora modules. in this case sql-lora and sql-lora2 point to the same underlying lora, but this need not be the case. lora config values take the same values they do in EngineArgs

no work has been done here to scope client permissions to specific models
This commit is contained in:
jvmncs
2024-02-17 12:00:48 -08:00
committed by GitHub
parent 185b2c29e2
commit 8f36444c4f
7 changed files with 200 additions and 27 deletions
+40 -1
View File
@@ -1,4 +1,5 @@
import asyncio
from dataclasses import dataclass
from http import HTTPStatus
from typing import Dict, List, Optional, Union
from vllm.logger import init_logger
@@ -9,15 +10,35 @@ from vllm.entrypoints.openai.protocol import (CompletionRequest,
ErrorResponse, LogProbs,
ModelCard, ModelList,
ModelPermission)
from vllm.lora.request import LoRARequest
logger = init_logger(__name__)
@dataclass
class LoRA:
name: str
local_path: str
class OpenAIServing:
def __init__(self, engine: AsyncLLMEngine, served_model: str):
def __init__(self,
engine: AsyncLLMEngine,
served_model: str,
lora_modules=Optional[List[LoRA]]):
self.engine = engine
self.served_model = served_model
if lora_modules is None:
self.lora_requests = []
else:
self.lora_requests = [
LoRARequest(
lora_name=lora.name,
lora_int_id=i,
lora_local_path=lora.local_path,
) for i, lora in enumerate(lora_modules, start=1)
]
self.max_model_len = 0
self.tokenizer = None
@@ -50,6 +71,13 @@ class OpenAIServing:
root=self.served_model,
permission=[ModelPermission()])
]
lora_cards = [
ModelCard(id=lora.lora_name,
root=self.served_model,
permission=[ModelPermission()])
for lora in self.lora_requests
]
model_cards.extend(lora_cards)
return ModelList(data=model_cards)
def _create_logprobs(
@@ -99,11 +127,22 @@ class OpenAIServing:
async def _check_model(self, request) -> Optional[ErrorResponse]:
if request.model == self.served_model:
return
if request.model in [lora.lora_name for lora in self.lora_requests]:
return
return self.create_error_response(
message=f"The model `{request.model}` does not exist.",
err_type="NotFoundError",
status_code=HTTPStatus.NOT_FOUND)
def _maybe_get_lora(self, request) -> Optional[LoRARequest]:
if request.model == self.served_model:
return
for lora in self.lora_requests:
if request.model == lora.lora_name:
return lora
# if _check_model has been called earlier, this will be unreachable
raise ValueError("The model `{request.model}` does not exist.")
def _validate_prompt_and_tokenize(
self,
request: Union[ChatCompletionRequest, CompletionRequest],