[Hack] Add pd-disaggregation decode polling interval (#10411)

python/sglang/srt/server_args.py

@@ -394,6 +394,9 @@ class ServerArgs:
     disaggregation_ib_device: Optional[str] = None
     num_reserved_decode_tokens: int = 512  # used for decode kv cache offload in PD
 
+    # FIXME: hack to reduce ITL when decode bs is small
+    disaggregation_decode_polling_interval: int = 1
+
     # For model weight update
     custom_weight_loader: Optional[List[str]] = None
     weight_loader_disable_mmap: bool = False
@@ -2245,6 +2248,12 @@ class ServerArgs:
             default=ServerArgs.num_reserved_decode_tokens,
             help="Number of decode tokens that will have memory reserved when adding new request to the running batch.",
         )
+        parser.add_argument(
+            "--disaggregation-decode-polling-interval",
+            type=int,
+            default=ServerArgs.disaggregation_decode_polling_interval,
+            help="The interval to poll requests in decode server. Can be set to >1 to reduce the overhead of this.",
+        )
 
         # Custom weight loader
         parser.add_argument(