services:
  qwen38-b70:
    build: .
    image: qwen38-b70:latest
    container_name: qwen38-b70
    restart: unless-stopped
    devices:
      - /dev/dri:/dev/dri
    volumes:
      # The Intel level-zero GPU driver enumerates via by-path; --device does
      # not recreate it, it must be mounted explicitly.
      - /dev/dri/by-path:/dev/dri/by-path:ro
      - /home/sero/models/q38-pinned:/models
      - /home/sero/qwen38-b70/entrypoint.sh:/entrypoint.sh:ro
    environment:
      - MODELS_DIR=/models
      - GPU_COUNT=2        # 1 or 2 Arc Pro B70s (1 GPU: no TP, higher prefill)
      - CTX_SIZE_OVERRIDE= # blank = 262144 (2 GPU) / 131072 (1 GPU)
      - PARALLEL=4
      - BATCH=8192
      - UBATCH=8192
      - ENABLE_MTP=1       # 1 = speculative decoding via mtp-Q4_0 draft
      - SPEC_DRAFT_N_MAX=5  # max draft tokens per step (5 = sweet spot)
      - THREADS=16          # CPU threads for target model
      - DRAFT_THREADS=16    # CPU threads for draft model
      - DRAFT_NGL=0         # 0 = CPU draft (runs in parallel with GPU target, no contention)
      - ENABLE_VISION=0    # 1 = image input via mmproj (CPU encoder; GPU offload hangs xe)
    ipc: host
    shm_size: 8g
    ports:
      - "8010:8010"