dev-rcfa-memory/be-ircfa/docker-compose.yaml

services:
  ircfa-app:
    build:
      context: .
      dockerfile: Dockerfile
    container_name: digitaltwin_ircfa
    working_dir: /app
    restart: always
    ports:
      - "3025:3021"
    environment:
      - ENV=development
      - LOG_LEVEL=INFO
      - PORT=3021
      - HOST=0.0.0.0
      - DATABASE_HOSTNAME=192.168.1.85
      - DATABASE_PORT=5432
      - DATABASE_CREDENTIAL_USER=postgres
      - DATABASE_CREDENTIAL_PASSWORD=postgres
      - DATABASE_NAME=digital_twin
      - COLLECTOR_HOSTNAME=192.168.1.86
      - COLLECTOR_PORT=5432
      - COLLECTOR_CREDENTIAL_USER=postgres
      - COLLECTOR_CREDENTIAL_PASSWORD=postgres
      - COLLECTOR_NAME=digital_twin
      - PFI_API_URL=http://192.168.1.82:8000/pfi/api/v1
      - RP_API_URL=http://192.168.1.82:8000/reliability
      - PLANT_STATUS_URL=http://192.168.1.82:8000/monitoring/plant-status
      - WINDOWS_AEROS_BASE_URL=http://192.168.1.87:8800
      - MODEL_SERVER_URL="https://dev-new-kong.air.id/dev-poc-its/v1"
      - OLLAMA_BASE_URL=http://192.168.1.100:11434
      - OLLAMA_MODEL_NAME=qwen3:8b
      - VLLM_BASE_URL=http://192.168.1.100:8020/v1
      - VLLM_MODEL_NAME=/models/qwen_6_std
    command: make run
    volumes:
      - ./src/uploads:/app/src/uploads


  kong:
    image: kong/kong-gateway:3.7.1.2
    container_name: kong-dbless
    restart: always
    environment:
      - KONG_DATABASE=off
      - KONG_DECLARATIVE_CONFIG=/kong/declarative/kong.yml
      - KONG_PROXY_ACCESS_LOG=/dev/stdout
      - KONG_ADMIN_ACCESS_LOG=/dev/stdout
      - KONG_PROXY_ERROR_LOG=/dev/stderr
      - KONG_ADMIN_ERROR_LOG=/dev/stderr
      - KONG_ADMIN_LISTEN=0.0.0.0:8001
      - KONG_ADMIN_GUI_URL=http://192.168.1.100:8002
      - KONG_LICENSE_DATA
      - KONG_NGINX_PROXY_PROXY_REQUEST_BUFFERING=off
    ports:
      - "8000:8000"
      - "8443:8443"
      - "8001:8001"
      - "8444:8444"
      - "8010:8002"
      - "8445:8445"
      - "8003:8003"
      - "8004:8004"
    volumes:
      - .:/kong/declarative # Volume mount for declarative config

  # Use the code below to apply the vllm service
  # --reasoning_parser deepseek-r1
  vllm:
    image: vllm/vllm-openai:latest
    runtime: nvidia
    environment:
      - CUDA_VISIBLE_DEVICES=0,1
      - VLLM_LOG_LEVEL=INFO
    volumes:
      - /mnt/data_backup/models:/models
    ports:
      - "8020:8000"
    ipc: host
    command: >
      --model "Qwen/Qwen3-30B-A3B-Instruct-2507-FP8"
      --dtype float16
      --tensor-parallel-size 2
      --max-model-len 16382
      --max-num-batched-tokens 3072
      --max-num-seqs 64
      --enforce-eager
      --trust-remote-code true
      --enable_prefix_caching true      

  llama-cpp:
    image: ghcr.io/ggml-org/llama.cpp:server-cuda
    container_name: llama-cpp-server
    restart: unless-stopped
    mem_limit: 40g
    memswap_limit: 40g
    oom_kill_disable: false
    oom_score_adj: 100
    ports:
      - "8087:8087"
    volumes:
      - /mnt/data_backup/model-qwen:/data:ro
    # Using environment variables for easy model swapping and parameter tuning
    command: >
      -m /data/Qwen3-30B-A3B-Q4_K_M.gguf
      --reasoning off
      -ngl ${LLAMA_CPP_N_GPU_LAYERS:-99}
      -c ${LLAMA_CPP_CTX_SIZE:-131072}
      -t ${LLAMA_CPP_THREADS:-8}
      -b ${LLAMA_CPP_BATCH_SIZE:-2048}
      -ub ${LLAMA_CPP_UBATCH_SIZE:-512}
      --temp 0
      --top-k 20
      --top-p 0.95
      --min-p 0
      -fa on
      -sm row
      --rope-scaling ${LLAMA_CPP_ROPE_SCALING:-yarn}
      --rope-scale ${LLAMA_CPP_ROPE_SCALE:-4}
      --yarn-orig-ctx ${LLAMA_CPP_YARN_ORIG_CTX:-32768}
      --no-context-shift
      --host 0.0.0.0
      --port 8087      
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8087/health"]
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 10s


    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              capabilities: [gpu]
              device_ids: ['all']

  vllm-gemma:
    image: vllm/vllm-openai:latest
    runtime: nvidia
    container_name: vllm-gemma-4
    environment:
      - CUDA_VISIBLE_DEVICES=0,1
      - VLLM_LOG_LEVEL=INFO
      - HF_TOKEN=${HF_TOKEN}
    volumes:
      - /mnt/data_backup/models:/root/.cache/huggingface
    ports:
      - "8021:8000"
    ipc: host
    command: >
      --model "unsloth/gemma-4-E2B-it"
      --tensor-parallel-size 2
      --max-model-len 8192
      --gpu-memory-utilization 0.90      

  llamacpp-gemma:
    image: ghcr.io/ggml-org/llama.cpp:server-cuda
    container_name: llamacpp-gemma-4
    restart: unless-stopped
    mem_limit: 40g
    memswap_limit: 40g
    oom_kill_disable: false
    oom_score_adj: 100
    ports:
      - "8088:8087"
    volumes:
      - /mnt/data_backup/models:/data:ro
    command: >
      -m /data/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf
      --mmproj /data/mmproj-BF16.gguf
      --reasoning off
      -ngl ${LLAMA_CPP_N_GPU_LAYERS:-99}
      -c ${LLAMA_CPP_CTX_SIZE:-196608}
      -t ${LLAMA_CPP_THREADS:-8}
      -b ${LLAMA_CPP_BATCH_SIZE:-2048}
      -ub ${LLAMA_CPP_UBATCH_SIZE:-512}
      --temp 0
      --top-k 20
      --top-p 0.95
      --min-p 0
      -fa on
      -sm layer
      --rope-scaling yarn
      --rope-scale 1.5
      --yarn-orig-ctx 131072
      --no-context-shift
      --host 0.0.0.0
      --port 8087      
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8087/health"]
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 10s

    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              capabilities: [gpu]
              device_ids: ['all']

  # --- ALTERNATIVE GEMMA 4 MODELS FOR TESTING AND COMPARISON ---
  # Uncomment the service you wish to test. Ensure you have downloaded the corresponding GGUF file.

  # 1. Gemma 4 E4B (4B parameters, ~7.6GB Q8_0, highly efficient, runs on GPU 0)
  # llamacpp-gemma-e4b:
  #   image: ghcr.io/ggml-org/llama.cpp:server-cuda
  #   container_name: llamacpp-gemma-e4b
  #   restart: unless-stopped
  #   ports:
  #     - "8088:8087"  # Maps to 8088 (same port as standard gemma service)
  #   volumes:
  #     - /mnt/data_backup/models:/data:ro
  #   environment:
  #     - CUDA_VISIBLE_DEVICES=0
  #   command: >
  #     -m /data/gemma-4-E4B-it-Q8_0.gguf
  #     --host 0.0.0.0
  #     --port 8087
  #     -ngl 99
  #     -c 8192
  #     --parallel 1
  #     --temp 0.15
  #     --reasoning off
  #   healthcheck:
  #     test: ["CMD", "curl", "-f", "http://localhost:8087/health"]
  #     interval: 30s
  #     timeout: 5s
  #     retries: 3
  #     start_period: 10s
  #   deploy:
  #     resources:
  #       reservations:
  #         devices:
  #           - driver: nvidia
  #             capabilities: [gpu]
  #             device_ids: ['0']

  # 2. Gemma 4 26B (26B parameters, ~15.8GB UD-Q4_K_M, runs on GPU 1 to balance load)
  # llamacpp-gemma-26b:
  #   image: ghcr.io/ggml-org/llama.cpp:server-cuda
  #   container_name: llamacpp-gemma-26b
  #   restart: unless-stopped
  #   ports:
  #     - "8089:8087"  # Runs on port 8089
  #   volumes:
  #     - /mnt/data_backup/models:/data:ro
  #   environment:
  #     - CUDA_VISIBLE_DEVICES=1
  #   command: >
  #     -m /data/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf
  #     --host 0.0.0.0
  #     --port 8087
      # -ngl ${LLAMA_CPP_N_GPU_LAYERS:-99}
      # -c ${LLAMA_CPP_CTX_SIZE:-131072}
      # -t ${LLAMA_CPP_THREADS:-8}
      # -b ${LLAMA_CPP_BATCH_SIZE:-2048}
      # -ub ${LLAMA_CPP_UBATCH_SIZE:-512}
  #     --parallel 1
  #     --temp 0.15
  #     -sm layer
  #     --reasoning off
  #   healthcheck:
  #     test: ["CMD", "curl", "-f", "http://localhost:8087/health"]
  #     interval: 30s
  #     timeout: 5s
  #     retries: 3
  #     start_period: 10s
  #   deploy:
  #     resources:
  #       reservations:
  #         devices:
  #           - driver: nvidia
  #             capabilities: [gpu]
  #             device_ids: ['1']

  # 3. Gemma 4 31B (31B parameters, ~17GB UD-Q4_K_M, runs on GPU 1)
  # llamacpp-gemma-31b:
  #   image: ghcr.io/ggml-org/llama.cpp:server-cuda
  #   container_name: llamacpp-gemma-31b
  #   restart: unless-stopped
  #   ports:
  #     - "8090:8087"  # Runs on port 8090
  #   volumes:
  #     - /mnt/data_backup/models:/data:ro
  #   environment:
  #     - CUDA_VISIBLE_DEVICES=1
  #   command: >
  #     -m /data/gemma-4-31B-it-Q4_K_M.gguf
  #     --host 0.0.0.0
  #     --port 8087
  # -ngl ${LLAMA_CPP_N_GPU_LAYERS:-99}
  # -c ${LLAMA_CPP_CTX_SIZE:-131072}
  # -t ${LLAMA_CPP_THREADS:-8}
  # -b ${LLAMA_CPP_BATCH_SIZE:-2048}
  # -ub ${LLAMA_CPP_UBATCH_SIZE:-512}
  #     --parallel 1
  #     --temp 0.15
  #     --reasoning off
  #     --rope-scaling ${LLAMA_CPP_ROPE_SCALING:-yarn}
  #     --rope-scale ${LLAMA_CPP_ROPE_SCALE:-4}
  #     --yarn-orig-ctx ${LLAMA_CPP_YARN_ORIG_CTX:-32768}
  #   healthcheck:
  #     test: ["CMD", "curl", "-f", "http://localhost:8087/health"]
  #     interval: 30s
  #     timeout: 5s
  #     retries: 3
  #     start_period: 10s
  #   deploy:
  #     resources:
  #       reservations:
  #         devices:
  #           - driver: nvidia
  #             capabilities: [gpu]
  #             device_ids: ['1']