Reduce scheduler recv requests overhead (#8947)

python/sglang/srt/server_args.py

@@ -249,6 +249,7 @@ class ServerArgs:
     enable_return_hidden_states: bool = False
     enable_triton_kernel_moe: bool = False
     enable_flashinfer_mxfp4_moe: bool = False
+    scheduler_recv_interval: int = 1
 
     # Debug tensor dumps
     debug_tensor_dump_output_folder: Optional[str] = None
@@ -1845,6 +1846,12 @@ class ServerArgs:
             action="store_true",
             help="Enable FlashInfer MXFP4 MoE backend for modelopt_fp4 quant on Blackwell.",
         )
+        parser.add_argument(
+            "--scheduler-recv-interval",
+            type=int,
+            default=ServerArgs.scheduler_recv_interval,
+            help="The interval to poll requests in scheduler. Can be set to >1 to reduce the overhead of this.",
+        )
 
         # Debug tensor dumps
         parser.add_argument(