저장

.gitignore

@@ -0,0 +1,3 @@
+model_services/paddle_vl/output
+data
+model_services/dotc.ocr/dots

docker-compose.yml

@@ -1,42 +1,37 @@
-services:
+1version: '3.8'
-  deepseek_ocr_vllm:
-    build:
-      context: .
-      dockerfile: Dockerfile.vllm
-    image: deepseek-ocr-vllm:cu126
-    container_name: deepseek_ocr_vllm
-    working_dir: /workspace
-    volumes:
-      - ./DeepSeek-OCR-master/DeepSeek-OCR-vllm:/workspace
-    gpus: all
-    shm_size: "8g"
-    ipc: "host"
-    environment:
-      - HF_HOME=/workspace/.cache/huggingface
-      - CUDA_HOME=/usr/local/cuda
-      - LD_LIBRARY_PATH=/usr/local/cuda/lib64:${LD_LIBRARY_PATH}
-      - PIP_DISABLE_PIP_VERSION_CHECK=1
-      - PYTHONUNBUFFERED=1
-    tty: true
-    entrypoint: ["/bin/bash"]
 
-  # deepseek_ocr_hf:
+services:
-  #   build:
+  api_gateway:
-  #     context: .
+    build:
-  #     dockerfile: Dockerfile.hf
+      context: ./api_gateway
-  #   image: deepseek-ocr-hf:cu126
+    ports:
-  #   container_name: deepseek_ocr_hf
+      - "80:8000" # 호스트의 80번 포트를 게이트웨이의 8000번 포트로 연결
-  #   working_dir: /workspace
+    networks:
-  #   volumes:
+      - vlm_network
-  #     - ./DeepSeek-OCR-hf:/workspace
+    depends_on:
-  #   gpus: all
+      - deepseek_ocr # deepseek_ocr 서비스가 시작된 후에 게이트웨이를 시작
-  #   shm_size: "8g"
+    restart: always
-  #   ipc: "host"
+
-  #   environment:
+  deepseek_ocr:
-  #     - HF_HOME=/workspace/.cache/huggingface
+    build:
-  #     - CUDA_HOME=/usr/local/cuda
+      context: ./model_services/deepseek_ocr
-  #     - LD_LIBRARY_PATH=/usr/local/cuda/lib64:${LD_LIBRARY_PATH}
+    # deploy 키를 사용하여 GPU 리소스를 요청합니다.
-  #     - PIP_DISABLE_PIP_VERSION_CHECK=1
+    # 이 설정은 docker-compose up --build 대신 docker stack deploy 또는 docker compose up --deploy와 함께 사용할 때 공식적으로 지원됩니다.
-  #     - PYTHONUNBUFFERED=1
+    # 일반적인 docker-compose up에서도 동작하는 경우가 많습니다.
-  #   tty: true
+    deploy:
-  #   entrypoint: ["/bin/bash"]
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1 # 1개의 GPU를 할당
+              capabilities: [gpu]
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=all # 컨테이너가 모든 GPU를 볼 수 있도록 설정
+      - MODEL_PATH=deepseek-ai/deepseek-vl-7b-base # 사용할 모델 지정 (필요시 수정)
+    networks:
+      - vlm_network
+    restart: on-failure
+
+networks:
+  vlm_network:
+    driver: bridge

model_services/paddle_vl/main.py

@@ -0,0 +1,12 @@
+from paddleocr import PaddleOCRVL
+
+pipeline = PaddleOCRVL(
+    vl_rec_backend="vllm-server", vl_rec_server_url="http://127.0.0.1:8118/v1"
+)
+output = pipeline.predict(
+    "/home/jackjack/test/deepseek_ocr/data/20250708092450-131-769-592.jpg"
+)
+for res in output:
+    res.print()
+    res.save_to_json(save_path="output")
+    res.save_to_markdown(save_path="output")