Keep image generic; move startup settings into run.sh/compose

docker-compose.yml

@@ -6,6 +6,13 @@ services:
     image: llama4-bi150:${TAG:-2026-10-26}
     container_name: llama4
 
+    # 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/llama4-entrypoint.sh"]
+    working_dir: /opt/llama4
+    expose:
+      - "8091"
+      - "8092"
+
     # vLLM 多卡 NCCL / 端口直通
     network_mode: host
     ipc: host
@@ -21,7 +28,7 @@ services:
     environment:
       MODEL_PATH: /models/llama4
       AUTO_DETECT: "${AUTO_DETECT:-1}"
-      HOST: 0.0.0.0
+      HOST: "${HOST:-0.0.0.0}"
 
       # 默认暴露全部 16 张卡；entrypoint 会根据模型选择实际 TP*PP。
       CUDA_VISIBLE_DEVICES: "${GPUS:-0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15}"
@@ -37,6 +44,7 @@ services:
       MAX_NUM_SEQS: "${MAX_NUM_SEQS:-1}"
       GPU_MEM_UTIL: "${GPU_MEM_UTIL:-0.90}"
       ENFORCE_EAGER: "${ENFORCE_EAGER:-1}"
+      PYTHONUNBUFFERED: "1"
       EXTRA_ARGS: "${EXTRA_ARGS:-}"
 
     ulimits: