Unified Gemma 4 family image for Gemma-4-31B and GEV-26B-Decide

- detect_model.py picks TP/port/served name/eager/util from config.json
- validate TP against the model config (dense global KV heads, MoE k % 32)
- one apply_patch.py covers both variants, idempotent
- Dockerfile stays generic: no ENV/WORKDIR/EXPOSE/ENTRYPOINT
- run.sh/verify.sh follow the llama4-bi150 layout

docker-compose.yml

@@ -0,0 +1,57 @@
+services:
+  gemma4:
+    build:
+      context: .
+      dockerfile: Dockerfile
+    image: gemma4-bi150:${TAG:-2026-10-27}
+    container_name: gemma4
+
+    # 镜像本身不设置启动命令，启动方式全部由这里决定。
+    entrypoint: ["/usr/local/bin/gemma4-entrypoint.sh"]
+    working_dir: /opt/gemma4
+    expose:
+      - "8094"
+      - "8095"
+
+    # vLLM 多卡 NCCL / 端口直通
+    network_mode: host
+    ipc: host
+
+    # Iluvatar 运行时需要 privileged + /dev 绑定。
+    # 只给 --device=/dev/iluvatar* 会在 worker init 时报
+    #   cudaGetDeviceCount() ... Error 1001
+    privileged: true
+    pid: host
+
+    volumes:
+      - /dev:/dev
+      - ${MODEL_HOST_DIR:-/mnt/disk0/models/google/gemma-4-31B-it}:/models/gemma4:ro
+
+    environment:
+      MODEL_PATH: /models/gemma4
+      AUTO_DETECT: "${AUTO_DETECT:-1}"
+      HOST: "${HOST:-0.0.0.0}"
+
+      # 默认给 4 张（Gemma-4-31B 的 TP4）；GEV 只需要 2 张，
+      # entrypoint 按检测出的 TP 使用前 TP 张。
+      CUDA_VISIBLE_DEVICES: "${GPUS:-15,14,13,12}"
+
+      # 留空时由 entrypoint 根据 config.json 自动判断。
+      # 需要手动覆盖时直接设置同名环境变量即可。
+      TP: "${TP:-}"
+      PORT: "${PORT:-}"
+      SERVED_NAME: "${SERVED_NAME:-}"
+      MAX_NUM_SEQS: "${MAX_NUM_SEQS:-}"
+      GPU_MEM_UTIL: "${GPU_MEM_UTIL:-}"
+      ENFORCE_EAGER: "${ENFORCE_EAGER:-}"
+
+      MAX_MODEL_LEN: "${MAX_MODEL_LEN:-8192}"
+      MAX_IMAGES: "${MAX_IMAGES:-4}"
+      PYTHONUNBUFFERED: "1"
+      EXTRA_ARGS: "${EXTRA_ARGS:-}"
+
+    ulimits:
+      memlock: -1
+      stack: 67108864
+
+    restart: unless-stopped