didi-lot1-ai/ai_platform/local_gpu_stack/docker-compose.yml

# Local GPU inference stack — 10.11.10.18 (2x H200 NVL)
# GPU 0: Qwen3.5-35B-A3B (LLM) + Whisper (audio)
# GPU 1: BusterX++ (deepfake video) + bge-m3 (embeddings) + bge-reranker-v2-m3 (rerank)
# All services on the shared `didi-network` so the dashboard can probe them by name.

networks:
  didi-network:
    external: true

services:
  # ===================== vLLM model servers =====================

  # BusterX++ — served OFFLINE from the on-disk HF cache (no download)
  vllm-buster:
    container_name: didiAI-video-vllm-buster
    # qwen3_5 image (vLLM 0.16 / transformers 4.57) — BusterX++ config was saved
    # with transformers 5.x and the older 0.8.5 image can't parse its nested text_config.
    image: vllm/vllm-openai:qwen3_5
    networks: [didi-network]
    ports: ["54500:54500"]
    volumes:
      - ${MODELS_DIR}:/root/.cache/huggingface/hub:ro
    environment:
      - HF_HOME=/root/.cache/huggingface
      - HF_HUB_OFFLINE=1
      - TRANSFORMERS_OFFLINE=1
    command: >
      --model l8cv/BusterX_plusplus
      --host 0.0.0.0 --port 54500
      --served-model-name busterx
      --tensor-parallel-size 1
      --max-model-len 32768
      --gpu-memory-utilization 0.25
      --trust-remote-code
      --enable-prefix-caching
      --disable-log-requests
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ['1']
              capabilities: [gpu]
    restart: unless-stopped

  # Qwen3.5-35B-A3B — downloads ~72GB on first start
  vllm-qwen:
    container_name: didiAI-vllm-qwen3.5
    image: vllm/vllm-openai:qwen3_5
    networks: [didi-network]
    ports: ["14001:14001"]
    volumes:
      - ${MODELS_DIR}:/root/.cache/huggingface/hub:ro
    environment:
      - HF_HOME=/root/.cache/huggingface
      - HF_HUB_OFFLINE=1
      - TRANSFORMERS_OFFLINE=1
      - VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
    command: >
      --model Qwen/Qwen3.5-35B-A3B
      --host 0.0.0.0 --port 14001
      --served-model-name qwen3.5
      --tensor-parallel-size 1
      --max-model-len 32000
      --gpu-memory-utilization 0.65
      --trust-remote-code
      --disable-log-requests
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ['0']
              capabilities: [gpu]
    restart: unless-stopped

  # bge-m3 embeddings — downloads ~2GB on first start
  vllm-embeddings:
    container_name: didiAI-vllm-embeddings
    image: vllm/vllm-openai:v0.8.5
    networks: [didi-network]
    ports: ["14101:14101"]
    volumes:
      - ${MODELS_DIR}:/root/.cache/huggingface/hub:ro
    environment:
      - HF_HOME=/root/.cache/huggingface
      - HF_HUB_OFFLINE=1
      - TRANSFORMERS_OFFLINE=1
    command: >
      --model BAAI/bge-m3
      --host 0.0.0.0 --port 14101
      --served-model-name bge-m3
      --task embed
      --gpu-memory-utilization 0.10
      --trust-remote-code
      --disable-log-requests
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ['1']
              capabilities: [gpu]
    restart: unless-stopped

  # bge-reranker-v2-m3 — downloads ~2GB on first start
  vllm-rerank:
    container_name: didiAI-vllm-rerank
    image: vllm/vllm-openai:v0.8.5
    networks: [didi-network]
    ports: ["14201:14201"]
    volumes:
      - ${MODELS_DIR}:/root/.cache/huggingface/hub:ro
    environment:
      - HF_HOME=/root/.cache/huggingface
      - HF_HUB_OFFLINE=1
      - TRANSFORMERS_OFFLINE=1
    command: >
      --model BAAI/bge-reranker-v2-m3
      --host 0.0.0.0 --port 14201
      --served-model-name bge-reranker-v2-m3
      --task score
      --gpu-memory-utilization 0.10
      --trust-remote-code
      --disable-log-requests
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ['1']
              capabilities: [gpu]
    restart: unless-stopped

  # ===================== FastAPI app wrappers =====================

  llm-api:
    container_name: didiAI-llm-api
    image: didiai-llm-inference:audit
    networks: [didi-network]
    ports: ["14011:14011"]
    environment:
      - LLM_HOST=0.0.0.0
      - LLM_PORT=14011
      - LLM_DEFAULT_BACKEND=vllm
      - LLM_ENABLE_VLLM=true
      - LLM_ENABLE_LLAMACPP=false
      - LLM_VLLM_BASE_URL=http://didiAI-vllm-qwen3.5:14001
      - LLM_DEFAULT_MODEL=qwen3.5
      - LLM_EXTERNAL_URL=http://${HOST_IP}:14011
    restart: unless-stopped

  embeddings-api:
    container_name: didiAI-embeddings-api
    image: didiai-embeddings:audit
    networks: [didi-network]
    ports: ["14100:14100"]
    environment:
      - EMB_HOST=0.0.0.0
      - EMB_PORT=14100
      - EMB_DEFAULT_BACKEND=vllm
      - EMB_ENABLE_VLLM=true
      - EMB_ENABLE_LLAMACPP=false
      - EMB_VLLM_BASE_URL=http://didiAI-vllm-embeddings:14101
      - EMB_VLLM_MODEL=bge-m3
      - EMB_EXTERNAL_URL=http://${HOST_IP}:14100
    restart: unless-stopped

  rerank-api:
    container_name: didiAI-rerank-api
    image: didiai-rerank:audit
    networks: [didi-network]
    ports: ["14200:14200"]
    environment:
      - RERANK_HOST=0.0.0.0
      - RERANK_PORT=14200
      - RERANK_DEFAULT_BACKEND=vllm
      - RERANK_ENABLE_VLLM=true
      - RERANK_ENABLE_LLAMACPP=false
      - RERANK_VLLM_BASE_URL=http://didiAI-vllm-rerank:14201
      - RERANK_VLLM_MODEL=bge-reranker-v2-m3
      - RERANK_EXTERNAL_URL=http://${HOST_IP}:14200
    restart: unless-stopped

  video-api:
    container_name: didiAI-video-api
    image: didiai-video-analysis:audit
    networks: [didi-network]
    ports: ["54600:54600"]
    volumes:
      - ./runs:/app/runs
    environment:
      - VIDEO_ANALYSIS_HOST=0.0.0.0
      - VIDEO_ANALYSIS_PORT=54600
      - VIDEO_ANALYSIS_EXTERNAL_URL=http://${HOST_IP}:54600
      - VIDEO_ANALYSIS_VLLM_BASE_URL=http://didiAI-video-vllm-buster:54500
      - VIDEO_ANALYSIS_VLLM_MODEL=busterx
      - VIDEO_ANALYSIS_RUNS_DIR=/app/runs
      - VIDEO_ANALYSIS_SEMANTIC_LLM_BASE_URL=http://didiAI-llm-api:14011
      - VIDEO_ANALYSIS_SEMANTIC_AGGREGATION_MODEL=qwen3.5
      - VIDEO_ANALYSIS_SEMANTIC_LLM_API_KEY=none
    restart: unless-stopped

  # ===================== Non-GPU services =====================

  catalog-api:
    container_name: didiAI-catalog-api
    image: didiai-catalog-api:audit
    networks: [didi-network]
    environment:
      - CATALOG_HOST=0.0.0.0
      - CATALOG_PORT=11000
      - CATALOG_EXTERNAL_URL=http://${HOST_IP}
      - CATALOG_LLM_URL=http://didiAI-llm-api:14011
      - CATALOG_AUDIO_URL=http://didiAI-audio:54300
      - CATALOG_VIDEO_URL=http://didiAI-video-api:54600
      - CATALOG_WEB_URL=http://didiAI-web-api:51100
    restart: unless-stopped

  extractors:
    container_name: didiAI-extractors
    image: didiai-extractors:audit
    networks: [didi-network]
    ports: ["54400:54400"]
    volumes:
      # GLiNER (+ backbone mdeberta) servit offline din zona locala de modele
      - ${MODELS_DIR}:/models-hf/hub:ro
    environment:
      - EXTRACTORS_HOST=0.0.0.0
      - EXTRACTORS_PORT=54400
      - EXTRACTORS_LLM_GATEWAY_URL=http://didiAI-llm-api:14011
      - EXTRACTORS_DETECT_MODEL=/tmp/yolov8n.pt
      - YOLO_CONFIG_DIR=/tmp/ultralytics
      - HF_HOME=/models-hf
      - HF_HUB_OFFLINE=1
      - TRANSFORMERS_OFFLINE=1
    restart: unless-stopped

  forensic:
    container_name: didiAI-forensic
    image: didiai-forensic:audit
    networks: [didi-network]
    ports: ["8085:8080"]
    environment:
      - API_PORT=8080
    restart: unless-stopped

  # cloak — scraper SERP stealth (tier-3 fallback pentru web)
  cloak:
    container_name: didiAI-cloak
    image: didiai-cloak:audit
    networks: [didi-network]
    ports: ["8770:8770"]
    environment:
      - CLOAK_HOST=0.0.0.0
      - CLOAK_PORT=8770
    restart: unless-stopped

  # SearXNG metasearch (motorul de cautare liber folosit de web-api)
  searxng:
    container_name: didiAI-web-searxng
    image: docker.io/searxng/searxng:latest
    networks: [didi-network]
    volumes:
      - ../../modules/web/deploy/metasearch/searxng:/etc/searxng:ro
    environment:
      - SEARXNG_BASE_URL=http://${HOST_IP}:8080/
    restart: unless-stopped

  web-api:
    container_name: didiAI-web-api
    image: didiai-web:audit
    networks: [didi-network]
    ports: ["51100:51100"]
    depends_on: [searxng]
    environment:
      - WEB_HOST=0.0.0.0
      - WEB_PORT=51100
      - WEB_EXTERNAL_URL=http://${HOST_IP}:51100
      - WEB_SEARXNG_BASE_URL=http://didiAI-web-searxng:8080
      - WEB_LLM_BASE_URL=http://didiAI-llm-api:14011
    restart: unless-stopped

  # Whisper STT — faster-whisper loads in-process on GPU 0
  audio-api:
    container_name: didiAI-audio
    image: didiai-audio:audit
    networks: [didi-network]
    ports: ["54300:54300"]
    volumes:
      # faster-whisper foloseste download_root=cache_dir -> models-- direct sub el (nu /hub)
      - ${MODELS_DIR}:/root/.cache/huggingface:ro
    environment:
      - AUDIO_HOST=0.0.0.0
      - AUDIO_PORT=54300
      - AUDIO_MODEL=large-v3-turbo
      - AUDIO_DEVICE=cuda
      - AUDIO_COMPUTE_TYPE=float16
      - AUDIO_CACHE_DIR=/root/.cache/huggingface
      - HF_HUB_OFFLINE=1
      - TRANSFORMERS_OFFLINE=1
      - AUDIO_EXTERNAL_URL=http://${HOST_IP}:54300
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ['0']
              capabilities: [gpu]
    restart: unless-stopped