[OPT]Skip the first delayer to maximize the BS of the decoding. (#19836)

python/sglang/srt/managers/prefill_delayer.py

@@ -66,6 +66,7 @@ class PrefillDelayer:
         self._metrics_collector = metrics_collector
 
         self._curr_state: Optional[_State] = None
+        self.skip_first_delayer = True
 
         assert (
             server_args.disaggregation_mode == "null"
@@ -146,8 +147,13 @@ class PrefillDelayer:
             if (
                 max_running_requests - global_running_batch.max().item()
                 < global_max_prefill_bs.max().item()
-                and not self.enable_dp_attention
             ):
+                # When the "max_decode_bs - running_bs < max_prefill_bs" condition is met,
+                # the first merge_batch causes the decoding to fail to reach the maximum batch size.
+                if self.skip_first_delayer:
+                    self.skip_first_delayer = False
+                    pass
+                else:
                     next_state = prev_state or _State()
                     next_state = next_state.bump_delayed_count()
                     return _NegotiateOutput(