[gRPC] Expose native pause status (#37488)

Co-authored-by: ishandhanani <82981111+ishandhanani@users.noreply.github.com>

python/sglang/srt/entrypoints/grpc_bridge.py

@@ -446,6 +446,9 @@ class RuntimeHandle:
             ServerStatus.UnHealthy,
         )
 
+    def get_is_ready(self) -> bool:
+        return self.tokenizer_manager.is_ready()
+
     def tokenize(self, text: str, add_special_tokens: bool = True) -> str:
         tokenizer = self.tokenizer_manager.tokenizer
         tokens = tokenizer.encode(text, add_special_tokens=add_special_tokens)

python/sglang/srt/entrypoints/http_server.py

@@ -659,6 +659,13 @@ async def validate_json_request(raw_request: Request):
 ##### Native API endpoints #####
 
 
+@app.get("/ready")
+async def ready() -> Response:
+    """Report whether the server is ready to receive new requests."""
+    status_code = 200 if _global_state.tokenizer_manager.is_ready() else 503
+    return Response(status_code=status_code)
+
+
 @app.get("/health")
 @app.get("/health_generate")
 async def health_generate(request: Request) -> Response:

python/sglang/srt/managers/tokenizer_manager.py

@@ -626,6 +626,14 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         # Subprocess liveness watchdog — set by Engine or http_server after construction
         self._subprocess_watchdog = None
 
+    def is_ready(self) -> bool:
+        """Return whether this server should receive new requests."""
+        return (
+            not self.is_pause
+            and not self.gracefully_exit
+            and self.server_status == ServerStatus.Up
+        )
+
     def init_request_logging_and_dumping(self):
         # TODO: Refactor and organize the log export code.
         # Request logging

python/sglang/srt/utils/auth.py

@@ -90,14 +90,19 @@ def decide_request_auth(
       it must be rejected (403) even if api_key is provided.
 
     NOTE :
-    - Health/metrics endpoints are always allowed (even when api_key/admin_api_key is set),
-      to support k8s/liveness/readiness and Prometheus scraping without embedding secrets.
+    - Health/readiness/metrics endpoints are always allowed (even when
+      api_key/admin_api_key is set), to support k8s probes and Prometheus
+      scraping without embedding secrets.
     - We match them by prefix to cover common variants like /health_generate.
     """
     if method == "OPTIONS":
         return AuthDecision(allowed=True)
 
-    if path.startswith("/health") or path.startswith("/metrics"):
+    if (
+        path.startswith("/health")
+        or path.startswith("/ready")
+        or path.startswith("/metrics")
+    ):
         return AuthDecision(allowed=True)
 
     def _check_bearer_token(