services:
  glm52:
    image: ${VERDICTAI_IMAGE:?set VERDICTAI_IMAGE to the accepted immutable image}
    restart: unless-stopped
    gpus: all
    ipc: host
    shm_size: "32g"
    ulimits:
      memlock: -1
      nofile: 1048576
    ports:
      - "${BIND_ADDRESS:-127.0.0.1}:${PORT:-8000}:8000"
    environment:
      CUDA_VISIBLE_DEVICES: "${CUDA_VISIBLE_DEVICES:-0,1,2,3,4,5,6,7}"
      VLLM_PP_LAYER_PARTITION: "9,10,10,10,10,10,10,9"
      VLLM_WORKER_MULTIPROC_METHOD: spawn
      VLLM_DEEP_GEMM_WARMUP: skip
    volumes:
      - "${MODEL_DIR:?set MODEL_DIR to this repository}:/model:ro"
      - "${CACHE_DIR:-./.runtime-cache}:/cache:rw"
    command:
      - vllm
      - serve
      - /model
      - --served-model-name
      - ${SERVED_MODEL_NAME:-GLM-5.2-SQG-W4A8}
      - --host
      - 0.0.0.0
      - --port
      - "8000"
      - --trust-remote-code
      - --tensor-parallel-size
      - "1"
      - --pipeline-parallel-size
      - "8"
      - --distributed-executor-backend
      - mp
      - --quantization
      - exl3
      - --dtype
      - bfloat16
      - --kv-cache-dtype
      - bfloat16
      - --load-format
      - safetensors
      - --attention-backend
      - FLASHMLA_SPARSE
      - --enforce-eager
      - --disable-custom-all-reduce
      - --gpu-memory-utilization
      - ${GPU_MEMORY_UTILIZATION:-0.90}
      - --max-model-len
      - ${MAX_MODEL_LEN:-262144}
      - --speculative-config
      - '{"method":"mtp","num_speculative_tokens":1}'
    healthcheck:
      test: ["CMD-SHELL", "curl -fsS http://127.0.0.1:8000/v1/models >/dev/null"]
      interval: 30s
      timeout: 10s
      retries: 60
      start_period: 20m