docker-compose.yml

services:
    web:
        build:
            context: ./web
            dockerfile: Dockerfile
        image: safelens-web
        ports:
            - "8000:8000"
        env_file:
            - ./web/.env
        environment:
            WHISPER_DEVICE: "cuda"
            WHISPER_COMPUTE_TYPE: "float16"
            WHISPER_MODEL: "large-v2"
            WHISPER_MAX_CONCURRENT: "1"
            WARMUP_ON_START: "true"
            NLTK_DATA: "/root/nltk_data"
            DATABASE_URL: "postgresql+psycopg://${WEB_DB_USER}:${WEB_DB_PASSWORD}@web_db:5432/${WEB_DB_NAME}"
            QWEN_VLLM_BASE_URL: "http://qwen_router:8080/v1"
            ANALYSIS_LLM_BACKEND: "http"
            ANALYSIS_LLM_HTTP_URL: "http://llama_router:8080/v1"
            QWEN_VLLM_MAX_INFLIGHT: "4"
        gpus:
            - driver: nvidia
              device_ids: ["3"]
              capabilities: [gpu]
        volumes:
            - videos_data:/app/videos
            - hf_cache:/root/.cache/huggingface
            - torch_cache:/root/.cache/torch
            - easyocr_cache:/root/.EasyOCR
            - nltk_data:/root/nltk_data
        depends_on:
            web_db:
                condition: service_healthy
        healthcheck:
            test: ["CMD-SHELL", "python3 -c \"import urllib.request; urllib.request.urlopen('http://localhost:8000/api/health/ready', timeout=3)\""]
            interval: 30s
            timeout: 5s
            retries: 10
        restart: unless-stopped
        profiles:
            - core

    web_db:
        image: postgres:17.5
        environment:
            POSTGRES_USER: ${WEB_DB_USER}
            POSTGRES_PASSWORD: ${WEB_DB_PASSWORD}
            POSTGRES_DB: ${WEB_DB_NAME}
        ports:
            - "5432:5432"
        volumes:
            - web_db_data:/var/lib/postgresql/data
        healthcheck:
            test: ["CMD-SHELL", "pg_isready -U socialaistudio -d harmful_moderation"]
            interval: 10s
            timeout: 5s
            retries: 5
        restart: unless-stopped
        profiles:
            - core

    auth_db:
        image: postgres:17.5
        environment:
            POSTGRES_USER: ${AUTH_DB_USER}
            POSTGRES_PASSWORD: ${AUTH_DB_PASSWORD}
            POSTGRES_DB: ${AUTH_DB_NAME}
        ports:
            - "5433:5432"
        volumes:
            - auth_db_data:/var/lib/postgresql/data
        healthcheck:
            test: ["CMD-SHELL", "pg_isready -U socialaistudio -d socialai_auth"]
            interval: 10s
            timeout: 5s
            retries: 5
        restart: unless-stopped
        profiles:
            - core

    vllm_llama:
        image: vllm/vllm-openai:nightly
        command:
            - "--host"
            - "0.0.0.0"
            - "--port"
            - "8192"
            - "--model"
            - "meta-llama/Meta-Llama-3-8B-Instruct"
            - "--enable-lora"
            - "--lora-modules"
            - "SafeLens/llama-3-8b=/models/lora"
        environment:
            HF_TOKEN: ${HF_TOKEN}
        gpus:
            - driver: nvidia
              device_ids: ["0"]
              capabilities: [gpu]
        ports:
            - "8192:8192"
        volumes:
            - ./web/lora_adapter:/models/lora:ro
            - vllm_cache:/root/.cache/huggingface
        restart: unless-stopped
        profiles:
            - gpu

    vllm_qwen:
        image: vllm/vllm-openai:nightly
        command:
            - "--host"
            - "0.0.0.0"
            - "--port"
            - "8193"
            - "--model"
            - "${VLLM_QWEN_MODEL:-Qwen/Qwen2.5-VL-7B-Instruct}"
            - "--trust-remote-code"
            - "--max-model-len"
            - "${VLLM_QWEN_MAX_MODEL_LEN:-8192}"
            - "--gpu-memory-utilization"
            - "${VLLM_QWEN_GPU_UTIL:-0.6}"
            - "--tensor-parallel-size"
            - "${VLLM_QWEN_TP:-1}"
        environment:
            HF_TOKEN: ${HF_TOKEN}
        gpus:
            - driver: nvidia
              device_ids: ["1"]
              capabilities: [gpu]
        ports:
            - "8193:8193"
        volumes:
            - vllm_cache:/root/.cache/huggingface
        healthcheck:
            test: ["CMD-SHELL", "python3 -c \"import urllib.request; urllib.request.urlopen('http://localhost:8193/v1/models', timeout=3)\""]
            interval: 30s
            timeout: 5s
            retries: 5
        restart: unless-stopped
        profiles:
            - gpu

    vllm_qwen_2:
        image: vllm/vllm-openai:nightly
        command:
            - "--host"
            - "0.0.0.0"
            - "--port"
            - "8194"
            - "--model"
            - "${VLLM_QWEN_MODEL:-Qwen/Qwen2.5-VL-7B-Instruct}"
            - "--trust-remote-code"
            - "--max-model-len"
            - "${VLLM_QWEN_MAX_MODEL_LEN:-8192}"
            - "--gpu-memory-utilization"
            - "${VLLM_QWEN_GPU_UTIL:-0.6}"
            - "--tensor-parallel-size"
            - "${VLLM_QWEN_TP:-1}"
        environment:
            HF_TOKEN: ${HF_TOKEN}
        gpus:
            - driver: nvidia
              device_ids: ["2"]
              capabilities: [gpu]
        ports:
            - "8194:8194"
        volumes:
            - vllm_cache:/root/.cache/huggingface
        healthcheck:
            test: ["CMD-SHELL", "python3 -c \"import urllib.request; urllib.request.urlopen('http://localhost:8194/v1/models', timeout=3)\""]
            interval: 30s
            timeout: 5s
            retries: 5
        restart: unless-stopped
        profiles:
            - gpu

    qwen_router:
        build:
            context: ./qwen-router
            dockerfile: Dockerfile
        image: safelens-qwen-router
        environment:
            # Override for remote/shared backends:
            #   export QWEN_ROUTER_BACKENDS="http://host:8193,http://host2:8194"
            LLM_ROUTER_BACKENDS: "${QWEN_ROUTER_BACKENDS:-http://vllm_qwen:8193,http://vllm_qwen_2:8194}"
            LLM_ROUTER_MAX_RETRIES: "1"
            LLM_ROUTER_COOLDOWN_SEC: "5"
            LLM_ROUTER_EWMA_ALPHA: "0.2"
            LLM_ROUTER_LATENCY_WEIGHT: "0"
        ports:
            - "8080:8080"
        depends_on:
            vllm_qwen:
                condition: service_started
        restart: unless-stopped
        profiles:
            - gpu

    llama_router:
        image: safelens-qwen-router
        environment:
            # Override for remote/shared backends:
            #   export LLAMA_ROUTER_BACKENDS="http://host:8192"
            LLM_ROUTER_BACKENDS: "${LLAMA_ROUTER_BACKENDS:-http://vllm_llama:8192}"
            LLM_ROUTER_MAX_RETRIES: "1"
            LLM_ROUTER_COOLDOWN_SEC: "5"
            LLM_ROUTER_EWMA_ALPHA: "0.2"
            LLM_ROUTER_LATENCY_WEIGHT: "0.001"
        ports:
            - "8081:8080"
        depends_on:
            vllm_llama:
                condition: service_started
        restart: unless-stopped
        profiles:
            - gpu

    cloudflared_web:
        image: cloudflare/cloudflared:latest
        command:
            - "tunnel"
            - "--config"
            - "/home/dylan/.cloudflared/harmful-moderation-config.yml"
            - "run"
        volumes:
            - ${HOME}/.cloudflared:/home/dylan/.cloudflared:ro
        depends_on:
            web:
                condition: service_started
        restart: unless-stopped
        profiles:
            - edge

    cloudflared_auth:
        image: cloudflare/cloudflared:latest
        command:
            - "tunnel"
            - "--config"
            - "/home/dylan/.cloudflared/socialai-auth-config.yml"
            - "run"
        volumes:
            - ${HOME}/.cloudflared:/home/dylan/.cloudflared:ro
        restart: unless-stopped
        profiles:
            - edge

volumes:
    videos_data:
    hf_cache:
    torch_cache:
    easyocr_cache:
    nltk_data:
    web_db_data:
    auth_db_data:
    vllm_cache: