Switch to bridge networking with fixed container port 80 and served name llm

- run.sh: bridge + -p ${HOST_PORT:-24396}:80, health check uses host port
- entrypoint.sh: PORT defaults to 80, SERVED_NAME defaults to llm
  detected port/served name become log hints only
- verify.sh: read host port from port bindings, expect model name llm
- docker-compose.yml: ports mapping instead of network_mode host
- README: update tables, commands and rationale

run.sh

@@ -16,7 +16,8 @@
 #   GPUS=12,13,14,15 ./run.sh /path/to/gemma4
 #   EXTRA_DOCKER_ARGS="-v /mnt/disk1:/data:ro" ./run.sh /path/to/model
 #   EXTRA_ARGS="--max-model-len 16384" ./run.sh /path/to/model
-#   TP=4 PORT=8099 SERVED_NAME=my-gemma4 ./run.sh /path/to/model
+#   HOST_PORT=24397 ./run.sh /path/to/model             # 换宿主机端口（默认 24396）
+#   TP=4 SERVED_NAME=llm ./run.sh /path/to/model
 #   BUILD=0 IMAGE=<ci-image> ./run.sh /path/to/model   # 使用已有镜像
 #   DRY_RUN=1 ./run.sh /path/to/model                  # 只打印 docker run 命令
 #
@@ -35,7 +36,8 @@ usage() {
 模型类型由 MODEL_DIR/config.json 自动判断，不需要 gemma4-31b / gev-26b-decide 参数。
 
 环境变量覆盖：
-  IMAGE NAME MODEL_HOST_DIR MODEL_MOUNT PORT SERVED_NAME TP GPUS
+  IMAGE NAME MODEL_HOST_DIR MODEL_MOUNT SERVED_NAME TP GPUS
+  HOST_PORT（宿主机端口，默认 24396；容器内固定 80）
   MAX_MODEL_LEN MAX_NUM_SEQS GPU_MEM_UTIL MAX_IMAGES ENFORCE_EAGER
   AUTO_DETECT EXTRA_ARGS EXTRA_DOCKER_ARGS BUILD=0/1 DRY_RUN=0/1
 EOF
@@ -81,9 +83,12 @@ MODEL_KIND="${MODEL_KIND:-${DETECTED_MODEL_KIND}}"
 IMAGE="${IMAGE:-gemma4-bi150:2026-10-27}"
 NAME="${NAME:-${MODEL_KIND}}"
 HOST="${HOST:-0.0.0.0}"
-PORT="${PORT:-${DETECTED_PORT}}"
+# 容器内端口固定 80；宿主机端口是 HOST_PORT，默认 24396（串行部署复用同一端口）。
+# PORT 保留为兼容别名，若显式给 PORT 则视作 HOST_PORT。
+CONTAINER_PORT=80
+HOST_PORT="${HOST_PORT:-${PORT:-24396}}"
 MODEL_MOUNT="${MODEL_MOUNT:-/models/gemma4}"
-SERVED_NAME="${SERVED_NAME:-${DETECTED_SERVED_NAME}}"
+SERVED_NAME="${SERVED_NAME:-llm}"
 TP="${TP:-${DETECTED_TP}}"
 MAX_MODEL_LEN="${MAX_MODEL_LEN:-8192}"
 MAX_NUM_SEQS="${MAX_NUM_SEQS:-${DETECTED_MAX_NUM_SEQS}}"
@@ -129,11 +134,11 @@ DOCKER_RUN=(
   docker run -d --name "${NAME}"
   --entrypoint /usr/local/bin/gemma4-entrypoint.sh
   --workdir /opt/gemma4
-  --expose "${PORT}"
+  --expose "${CONTAINER_PORT}"
   --privileged
   --pid host
-  --network host
   --ipc host
+  -p "${HOST_PORT}:${CONTAINER_PORT}"
   -v /dev:/dev
   -v "${MODEL_HOST_DIR}:${MODEL_MOUNT}:ro"
   --ulimit memlock=-1
@@ -144,7 +149,7 @@ DOCKER_RUN=(
   -e MODEL_KIND="${MODEL_KIND}"
   -e HOST="${HOST}"
   -e SERVED_NAME="${SERVED_NAME}"
-  -e PORT="${PORT}"
+  -e PORT="${CONTAINER_PORT}"
   -e TP="${TP}"
   -e MAX_MODEL_LEN="${MAX_MODEL_LEN}"
   -e MAX_NUM_SEQS="${MAX_NUM_SEQS}"
@@ -160,7 +165,9 @@ DOCKER_RUN=(
 
 echo "==> detected ${MODEL_KIND}: ${DETECTED_DESCRIPTION}"
 echo "    model : ${MODEL_HOST_DIR}"
-echo "    TP    : ${TP} (${GPU_COUNT} GPUs), port=${PORT}, name=${SERVED_NAME}"
+echo "    TP    : ${TP} (${GPU_COUNT} GPUs)"
+echo "    port  : 127.0.0.1:${HOST_PORT} -> container:${CONTAINER_PORT}"
+echo "    name  : ${SERVED_NAME}"
 echo "    GPUs  : ${GPUS}"
 echo "    eager : ${ENFORCE_EAGER}, util=${GPU_MEM_UTIL}, seqs=${MAX_NUM_SEQS}"
 
@@ -185,11 +192,11 @@ echo "==> run ${MODEL_KIND}"
 
 # 冷启动：Gemma-4-31B 约 59 GB、GEV 约 49 GB，首次读盘实测 5-8 分钟。
 HEALTH_TIMEOUT="${HEALTH_TIMEOUT:-900}"
-echo "==> wait for http://127.0.0.1:${PORT}/health (up to ${HEALTH_TIMEOUT}s)"
+echo "==> wait for http://127.0.0.1:${HOST_PORT}/health (up to ${HEALTH_TIMEOUT}s)"
 DEADLINE=$((SECONDS + HEALTH_TIMEOUT))
 while (( SECONDS < DEADLINE )); do
-  if curl -sf "http://127.0.0.1:${PORT}/health" >/dev/null 2>&1; then
-    echo "ready: http://127.0.0.1:${PORT}/v1/models"
+  if curl -sf "http://127.0.0.1:${HOST_PORT}/health" >/dev/null 2>&1; then
+    echo "ready: http://127.0.0.1:${HOST_PORT}/v1/models"
     exit 0
   fi
   sleep 5