레거시 모드

api.py

@@ -1,7 +1,19 @@
 import logging
 
+from config.env_setup import setup_environment
+
+# 환경 변수 설정을 최우선으로 호출
+setup_environment()
+
+# 로깅 기본 설정
+logging.basicConfig(
+    level=logging.INFO, format="%(asctime)s [%(levelname)s] %(name)s - %(message)s"
+)
+logger = logging.getLogger("startup")
+
 from fastapi import FastAPI
 from router import deepseek_router
+from services.ocr_engine import init_engine
 
 logging.basicConfig(
     level=logging.INFO, format="%(asctime)s [%(levelname)s] %(name)s - %(message)s"
@@ -14,6 +26,17 @@ app = FastAPI(
 )
 
 
+@app.on_event("startup")
+async def startup_event():
+    """FastAPI startup event handler."""
+    logging.info("Application startup...")
+    try:
+        await init_engine()
+        logging.info("vLLM engine initialized successfully.")
+    except Exception as e:
+        logging.error(f"vLLM engine init failed: {e}", exc_info=True)
+
+
 @app.get("/health/API", include_in_schema=False)
 async def health_check():
     return {"status": "API ok"}

services/ocr_engine.py

@@ -1,37 +1,45 @@
 import asyncio
 import io
-from typing import Union
+import logging
 
 import fitz
 from config.model_settings import CROP_MODE, MODEL_PATH, PROMPT
-from fastapi import UploadFile
 from PIL import Image
-from process.image_process import DeepseekOCRProcessor
 from vllm import AsyncLLMEngine, SamplingParams
 from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.model_executor.models.registry import ModelRegistry
 
 from services.deepseek_ocr import DeepseekOCRForCausalLM
+from services.process.image_process import DeepseekOCRProcessor
+
+logger = logging.getLogger(__name__)
 
 # --------------------------------------------------------------------------
 # 1. 모델 및 프로세서 초기화
 # --------------------------------------------------------------------------
 
-# VLLM이 커스텀 모델을 인식하도록 등록
 ModelRegistry.register_model("DeepseekOCRForCausalLM", DeepseekOCRForCausalLM)
+_engine = None
+
+
+async def init_engine():
+    """vLLM 엔진을 비동기적으로 초기화합니다."""
+    global _engine
+    if _engine is not None:
+        return
+
+    engine_args = AsyncEngineArgs(
+        model=MODEL_PATH,
+        hf_overrides={"architectures": ["DeepseekOCRForCausalLM"]},
+        block_size=256,
+        max_model_len=8192,
+        enforce_eager=False,
+        trust_remote_code=True,
+        tensor_parallel_size=1,
+        gpu_memory_utilization=0.75,
+    )
+    _engine = AsyncLLMEngine.from_engine_args(engine_args)
 
-# VLLM 비동기 엔진 설정
-engine_args = AsyncEngineArgs(
-    model=MODEL_PATH,
-    hf_overrides={"architectures": ["DeepseekOCRForCausalLM"]},
-    block_size=256,
-    max_model_len=8192,
-    enforce_eager=False,
-    trust_remote_code=True,
-    tensor_parallel_size=1,
-    gpu_memory_utilization=0.75,
-)
-engine = AsyncLLMEngine.from_engine_args(engine_args)
 
 # 샘플링 파라미터 설정
 sampling_params = SamplingParams(
@@ -47,8 +55,11 @@ processor = DeepseekOCRProcessor()
 # 2. 핵심 처리 함수
 # --------------------------------------------------------------------------
 
+
 async def _process_single_image(image: Image.Image) -> str:
     """단일 PIL 이미지를 받아 OCR을 수행하고 텍스트를 반환합니다."""
+    if _engine is None:
+        raise RuntimeError("vLLM engine not initialized yet")
     if "<image>" not in PROMPT:
         raise ValueError("프롬프트에 '<image>' 토큰이 없어 OCR을 수행할 수 없습니다.")
 
@@ -60,12 +71,13 @@ async def _process_single_image(image: Image.Image) -> str:
     request_id = f"request-{asyncio.get_running_loop().time()}"
 
     final_output = ""
-    async for request_output in engine.generate(request, sampling_params, request_id):
+    async for request_output in _engine.generate(request, sampling_params, request_id):
         if request_output.outputs:
             final_output = request_output.outputs[0].text
 
     return final_output
 
+
 def _pdf_to_images(pdf_bytes: bytes, dpi=144) -> list[Image.Image]:
     """PDF 바이트를 받아 페이지별 PIL 이미지 리스트를 반환합니다."""
     images = []
@@ -83,6 +95,7 @@ def _pdf_to_images(pdf_bytes: bytes, dpi=144) -> list[Image.Image]:
     pdf_document.close()
     return images
 
+
 async def process_document(file_bytes: bytes, content_type: str, filename: str) -> dict:
     """
     업로드된 파일(이미지 또는 PDF)을 처리하여 OCR 결과를 반환합니다.