services:
  vllm:
    build:
      context: .
      dockerfile: Dockerfile
    image: lfm25-8b-a1b-nvfp4
    container_name: vllm-lfm25
    runtime: nvidia
    restart: unless-stopped
    ports:
      - "8000:8000"
    environment:
      - MODEL_DIR=/model
      - PORT=8000
      - MAX_MODEL_LEN=${MAX_MODEL_LEN:-32768}
      - MAX_NUM_SEQS=${MAX_NUM_SEQS:-16}
      - GPU_MEM_UTIL=${GPU_MEM_UTIL:-0.90}
      - KV_CACHE_DTYPE=${KV_CACHE_DTYPE:-auto}
      - TP_SIZE=${TP_SIZE:-1}
      - REASONING_PARSER=${REASONING_PARSER:-deepseek_r1}
      - SERVED_MODEL_NAME=lfm25-8b-a1b
      - NVIDIA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-0}
      - TORCH_MATMUL_PRECISION=high
      - VLLM_USE_FLASHINFER_SAMPLER=1
    volumes:
      # point this at the model folder (defaults to this directory)
      - .:/model:ro
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1          # one Blackwell card is enough
              capabilities: [gpu]
    shm_size: 8g