spec: gate dp mlp sync with server args (#24177)

python/sglang/srt/managers/scheduler.py

@@ -2463,7 +2463,11 @@ class Scheduler(
             new_batch = self.get_new_batch_prefill()
 
         need_mlp_sync = self.require_mlp_sync
-        if need_mlp_sync and not self.spec_algorithm.is_none():
+        if (
+            need_mlp_sync
+            and not self.spec_algorithm.is_none()
+            and not self.server_args.speculative_skip_dp_mlp_sync
+        ):
             # NOTE: This branch makes sure prefill and decode batches will not be mixed when spec and dp-attn is enabled.
             # Before merging the new batch into running batch:
             # 1. All new batches are none -> need_mlp_sync remains true (sync is needed for decode batch).

python/sglang/srt/server_args.py

@@ -526,6 +526,7 @@ class ServerArgs:
     speculative_draft_model_quantization: Optional[str] = None
     speculative_adaptive: bool = False
     speculative_adaptive_config: Optional[str] = None
+    speculative_skip_dp_mlp_sync: bool = False
 
     # Speculative decoding (ngram)
     speculative_ngram_min_bfs_breadth: int = 1
@@ -3300,6 +3301,12 @@ class ServerArgs:
         if self.speculative_algorithm == "NEXTN":
             self.speculative_algorithm = "EAGLE"
 
+        if self.speculative_skip_dp_mlp_sync:
+            assert self.speculative_algorithm == "EAGLE", (
+                "--speculative-skip-dp-mlp-sync is only supported with "
+                f"speculative_algorithm == EAGLE, got {self.speculative_algorithm}."
+            )
+
         if self.speculative_algorithm == "DFLASH":
             if self.enable_dp_attention:
                 raise ValueError(
@@ -5497,6 +5504,13 @@ class ServerArgs:
             help="Path to a JSON config file for adaptive speculative decoding tuning knobs ",
             default=ServerArgs.speculative_adaptive_config,
         )
+        parser.add_argument(
+            "--speculative-skip-dp-mlp-sync",
+            action="store_true",
+            default=ServerArgs.speculative_skip_dp_mlp_sync,
+            help="Skip the extra MLP sync that the scheduler performs before merging a new batch "
+            "when speculative decoding + DP attention are both enabled.",
+        )
 
         # Multi-layer Eagle speculative decoding
         parser.add_argument(