version: '3.8'
services:
  qwen3.6-27b:
    image: cr.metax-tech.com/public-ai-release/maca/vllm-metax:0.19.0-maca.ai3.5.3.502-torch2.8-py312-ubuntu22.04-amd64
    container_name: qwen3.6-27b-w8a8-vllm #qwen3.6-27b-w8a8  Qwen3.6-27B-W8A8
    stdin_open: true
    tty: true
    restart: unless-stopped
    #network_mode: host
    devices:
      - "/dev/dri:/dev/dri"
      - "/dev/mxcd:/dev/mxcd"
      - "/dev/mem:/dev/mem"
    group_add:
      - "video"
    privileged: true
    security_opt:
      - "apparmor=unconfined"
      - "seccomp=unconfined"
    shm_size: '100gb'
    ulimits:
      memlock:
        soft: -1
        hard: -1
    ports:
      - "8004:30000"
    environment:
      - CUDA_VISIBLE_DEVICES=0,1
      - PYTHONUNBUFFERED=1  # 确保实时输出
      - MACA_SMALL_PAGESIZE_ENABLE=1
      - MACA_VLLM_ENABLE_MCTLASS_FUSED_MOE=1
      - MACA_VLLM_ENABLE_MCTLASS_PYTHON_API=1
    volumes:
      - "/usr/local/:/usr/local/"
      - "/pde_ai:/pde_ai"
      - "/opt/lq/models:/model:ro"
      - "~/.cache/huggingface:/root/.cache/huggingface"
      - "/opt/lq/deploy_models/logs:/var/log/vllm"  # 日志目录映射
      - "/opt/lq/deploy_models/bench_suite:/bench_suite" #脚本目录映射
    command: >
      sh -c "/opt/conda/bin/vllm serve /model/Qwen3.6-27B-W8A8 \
      --served-model-name Qwen3.6-27B-W8A8 \
      --host 0.0.0.0 \
      --port 30000 \
      --tensor-parallel-size 2 \
      --max-num-batched-tokens 4096 \
      --max-model-len 8192 \
      --reasoning-parser qwen3 \
      --enable-auto-tool-choice \
      --tool-call-parser qwen3_coder \ 
      --api-key sk-123456 \
      2>&1 | tee /var/log/vllm/qwen3.6-27b-w8a8-server.log"



  qwen3-embedding:
    image: vllm-metax:lq
    container_name: qwen3-embedding-vllm #qwen3-embedding
    stdin_open: true
    tty: true
    restart: unless-stopped
    #network_mode: host
    devices:
      - "/dev/dri:/dev/dri"
      - "/dev/mxcd:/dev/mxcd"
      - "/dev/mem:/dev/mem"
    group_add:
      - "video"
    privileged: true
    security_opt:
      - "apparmor=unconfined"
      - "seccomp=unconfined"
    shm_size: '100gb'
    ulimits:
      memlock:
        soft: -1
        hard: -1
    ports:
      - "9003:30000"
    environment:
      - CUDA_VISIBLE_DEVICES=2
      - PYTHONUNBUFFERED=1  # 确保实时输出
      - VLLM_TORCH_COMPILE=0
      - VLLM_DISABLE_TORCH_COMPILE=1
      - TORCH_EXTENSIONS_DIR=/tmp/torch_ext_$$
      - MAX_JOBS=1
    volumes:
      - "/usr/local/:/usr/local/"
      - "/pde_ai:/pde_ai"
      - "/opt/lq/models:/model:ro"
      - "~/.cache/huggingface:/root/.cache/huggingface"
      - "/opt/lq/deploy_models/logs:/var/log/vllm"  # 日志目录映射
      - "/opt/lq/deploy_models/bench_suite:/bench_suite" #脚本目录映射
    command: >
      sh -c "/opt/conda/bin/vllm serve /model/Qwen3-Embedding-8B  \
      --served-model-name Qwen3-Embedding-8B  \
      --task embedding \
      --host 0.0.0.0 \
      --port 30000 \
      --tensor-parallel-size 1 \
      --max-num-batched-tokens 4096 \
      --max-model-len 16384 \
      --gpu-memory-utilization 0.45 \
      --api-key sk-123456 \
      2>&1 | tee /var/log/vllm/qwen3-embedding-server.log"


  qwen3-reranker:
    image: vllm-metax:lq
    container_name: qwen3-reranker-vllm #qwen3-reranker
    stdin_open: true
    tty: true
    restart: unless-stopped
    #network_mode: host
    devices:
      - "/dev/dri:/dev/dri"
      - "/dev/mxcd:/dev/mxcd"
      - "/dev/mem:/dev/mem"
    group_add:
      - "video"
    privileged: true
    security_opt:
      - "apparmor=unconfined"
      - "seccomp=unconfined"
    shm_size: '100gb'
    ulimits:
      memlock:
        soft: -1
        hard: -1
    ports:
      - "9004:30000"
    environment:
      - CUDA_VISIBLE_DEVICES=3
      - PYTHONUNBUFFERED=1  # 确保实时输出
      - VLLM_TORCH_COMPILE=0
      - VLLM_DISABLE_TORCH_COMPILE=1
      - TORCH_EXTENSIONS_DIR=/tmp/torch_ext_$$
      - MAX_JOBS=1
    volumes:
      - "/usr/local/:/usr/local/"
      - "/pde_ai:/pde_ai"
      - "/opt/lq/models:/model:ro"
      - "~/.cache/huggingface:/root/.cache/huggingface"
      - "/opt/lq/deploy_models/logs:/var/log/vllm"  # 日志目录映射
      - "/opt/lq/deploy_models/bench_suite:/bench_suite" #脚本目录映射
    command: >
      sh -c "/opt/conda/bin/vllm serve /model/Qwen3-Reranker-8B  \
      --served-model-name Qwen3-Reranker-8B  \
      --task score \
      --host 0.0.0.0 \
      --port 30000 \
      --tensor-parallel-size 1 \
      --max-num-batched-tokens 4096 \
      --max-model-len 16384 \
      --gpu-memory-utilization 0.45 \
      --hf_overrides '{\"architectures\": [\"Qwen3ForSequenceClassification\"],\"classifier_from_token\": [\"no\", \"yes\"],\"is_original_qwen3_reranker\": true}' \
      --api-key sk-123456 \
      2>&1 | tee /var/log/vllm/qwen3-reranker-server.log"