[router][grpc] Add serve_grpc to launch_server and log id for HealthCheck (#11564)

python/sglang/launch_server.py

@@ -1,9 +1,9 @@
 """Launch the inference server."""
 
+import asyncio
 import os
 import sys
 
-from sglang.srt.entrypoints.http_server import launch_server
 from sglang.srt.server_args import prepare_server_args
 from sglang.srt.utils import kill_process_tree
 
@@ -11,6 +11,13 @@ if __name__ == "__main__":
     server_args = prepare_server_args(sys.argv[1:])
 
     try:
-        launch_server(server_args)
+        if server_args.grpc_mode:
+            from sglang.srt.entrypoints.grpc_server import serve_grpc
+
+            asyncio.run(serve_grpc(server_args))
+        else:
+            from sglang.srt.entrypoints.http_server import launch_server
+
+            launch_server(server_args)
     finally:
         kill_process_tree(os.getpid(), include_parent=False)

python/sglang/srt/entrypoints/grpc_server.py

@@ -22,8 +22,8 @@ from grpc_reflection.v1alpha import reflection
 
 import sglang
 from sglang.srt.disaggregation.utils import FAKE_BOOTSTRAP_HOST, DisaggregationMode
-from sglang.srt.entrypoints.grpc_request_manager import GrpcRequestManager
 from sglang.srt.grpc import sglang_scheduler_pb2, sglang_scheduler_pb2_grpc
+from sglang.srt.grpc.grpc_request_manager import GrpcRequestManager
 from sglang.srt.managers.data_parallel_controller import (
     run_data_parallel_controller_process,
 )
@@ -68,6 +68,8 @@ def _launch_scheduler_process_only(
     # Configure global environment
     configure_logger(server_args)
     server_args.check_server_args()
+    # Fix CUDA multiprocessing issues - must be called before any CUDA operations
+    mp.set_start_method("spawn", force=True)
 
     # Allocate ports for inter-process communications
     if port_args is None:
@@ -317,7 +319,8 @@ class SGLangSchedulerServicer(sglang_scheduler_pb2_grpc.SglangSchedulerServicer)
         Check the health of the inference server by sending a special request to generate one token.
         Similar to HTTP server's /health endpoint.
         """
-        logger.info("Receive health check request")
+        rid = f"HEALTH_CHECK_{time.time()}"
+        logger.info(f"Receive health check request: {rid}")
 
         if self.request_manager.gracefully_exit:
             logger.info(
@@ -328,7 +331,6 @@ class SGLangSchedulerServicer(sglang_scheduler_pb2_grpc.SglangSchedulerServicer)
             )
 
         # Create a special health check request
-        rid = f"HEALTH_CHECK_{time.time()}"
         sampling_params = SGLSamplingParams(max_new_tokens=1, temperature=0.0)
         sampling_params.normalize(tokenizer=None)
 
@@ -919,25 +921,3 @@ async def serve_grpc(
                     proc.join(timeout=1.0)
 
         logger.info("All scheduler processes terminated")
-
-
-def main():
-    """Main entry point for standalone gRPC server."""
-    # Fix CUDA multiprocessing issues - must be called before any CUDA operations
-    mp.set_start_method("spawn", force=True)
-
-    parser = argparse.ArgumentParser(description="SGLang Standalone gRPC Server")
-    ServerArgs.add_cli_args(parser)
-    args = parser.parse_args()
-    server_args = ServerArgs.from_cli_args(args)
-
-    # Run server
-    asyncio.run(
-        serve_grpc(
-            server_args=server_args,
-        )
-    )
-
-
-if __name__ == "__main__":
-    main()

python/sglang/srt/grpc/sglang_scheduler.proto

@@ -326,10 +326,7 @@ message EmbedError {
 // Management Operations
 // =====================
 
-message HealthCheckRequest {
-  // Input for health test generation (must be tokenized)
-  TokenizedInput tokenized = 1;
-}
+message HealthCheckRequest {}
 
 message HealthCheckResponse {
   bool healthy = 1;

python/sglang/srt/grpc/sglang_scheduler_pb2.pyi

@@ -320,10 +320,8 @@ class EmbedError(_message.Message):
     def __init__(self, message: _Optional[str] = ..., code: _Optional[str] = ..., details: _Optional[str] = ...) -> None: ...
 
 class HealthCheckRequest(_message.Message):
-    __slots__ = ("tokenized",)
-    TOKENIZED_FIELD_NUMBER: _ClassVar[int]
-    tokenized: TokenizedInput
-    def __init__(self, tokenized: _Optional[_Union[TokenizedInput, _Mapping]] = ...) -> None: ...
+    __slots__ = ()
+    def __init__(self) -> None: ...
 
 class HealthCheckResponse(_message.Message):
     __slots__ = ("healthy", "message")

python/sglang/srt/server_args.py

@@ -194,6 +194,7 @@ class ServerArgs:
     # HTTP server
     host: str = "127.0.0.1"
     port: int = 30000
+    grpc_mode: bool = False
     skip_server_warmup: bool = False
     warmups: Optional[str] = None
     nccl_port: Optional[int] = None
@@ -1516,6 +1517,11 @@ class ServerArgs:
             default=ServerArgs.port,
             help="The port of the HTTP server.",
         )
+        parser.add_argument(
+            "--grpc-mode",
+            action="store_true",
+            help="If set, use gRPC server instead of HTTP server.",
+        )
         parser.add_argument(
             "--skip-server-warmup",
             action="store_true",