.

Dockerfile

@@ -144,13 +144,14 @@ RUN ls -lh /tmp/wheels && \
     python3 -c "from torch.distributed import Backend; print('✅ Runtime torch distributed OK, GLOO =', Backend.GLOO)" && \
     rm -rf /tmp/wheels
 
-# # 安装运行时漏掉的依赖
-# RUN python3 -m pip install --no-cache-dir pydantic orjson psutil pyzmq pynvml transformers==4.48.3 uvicorn fastapi IPython aiohttp setproctitle uvloop sentencepiece triton
+# ✅ 安装 Prometheus client
+RUN python3 -m pip install --no-cache-dir prometheus_client
 
-# ✅ 离线安装全部依赖（包含所有运行时必需包）
-# RUN python3 -m pip install --no-cache-dir --no-deps /tmp/wheels/* && \
-#     python3 -c "from torch.distributed import Backend; print('✅ Runtime torch distributed OK, GLOO =', Backend.GLOO)" && \
-#     rm -rf /tmp/wheels
+# ✅ 设置多进程 metrics 收集目录（用于 MultiProcessCollector）
+ENV PROMETHEUS_MULTIPROC_DIR=/tmp/prometheus
+
+# ✅ 确保目录存在
+RUN mkdir -p /tmp/prometheus
 
 # ✅ 添加 Tini（推荐）
 ENV TINI_VERSION=v0.19.0
@@ -161,6 +162,8 @@ ENTRYPOINT ["/tini", "--"]
 # ---- 拷贝模型（路径可换） ----
 COPY ./Alibaba/Qwen3-8B /root/.cradle/Alibaba/Qwen3-8B
 
+HEALTHCHECK --interval=30s --timeout=2s --start-period=600s --retries=5 CMD curl -fs http://localhost:30000/health || exit 1
+
 # ---- 暴露端口 ----
 EXPOSE 30000
 
@@ -170,4 +173,5 @@ CMD ["python3", "-m", "sglang.launch_server", \
      "--port", "30000", \
      "--model-path", "/root/.cradle/Alibaba/Qwen3-8B/", \
      "--tp", "1", \
-     "--api-key", "token-abc123"]
+     "--api-key", "token-abc123", \
+     "--enable-metrics"]

sglang/python/sglang/srt/entrypoints/http_server.py

@@ -216,9 +216,12 @@ HEALTH_CHECK_TIMEOUT = int(os.getenv("SGLANG_HEALTH_CHECK_TIMEOUT", 20))
 
 
 @app.get("/health")
-async def health() -> Response:
-    """Check the health of the http server."""
-    return Response(status_code=200)
+async def health():
+    """Check the health of the http server and return version info."""
+    return {
+        "status": "ok",
+        "version": "v1.0.0"  # 这里写上你希望显示的版本号
+    }
 
 
 @app.get("/health_generate")

sglang/python/sglang/srt/utils.py

@@ -868,12 +868,22 @@ def set_ulimit(target_soft_limit=65535):
 def add_api_key_middleware(app, api_key: str):
     @app.middleware("http")
     async def authentication(request, call_next):
+        # OPTIONS 请求（CORS 预检）直接放行
         if request.method == "OPTIONS":
             return await call_next(request)
-        if request.url.path.startswith("/health"):
-            return await call_next(request)
-        if request.url.path.startswith("/metrics"):
+
+        # 明确列出无需鉴权的路径前缀
+        whitelist_prefixes = (
+            "/health",
+            "/metrics",
+            "/ping",
+            "/get_model_info",
+        )
+
+        if any(request.url.path.startswith(prefix) for prefix in whitelist_prefixes):
             return await call_next(request)
+
+        # Bearer Token 鉴权
         if request.headers.get("Authorization") != "Bearer " + api_key:
             return ORJSONResponse(content={"error": "Unauthorized"}, status_code=401)
         return await call_next(request)