[Spec] Consolidate the grammar sync decision into ScheduleBatch.grammar_needs_sync (#32353)

python/sglang/srt/managers/schedule_batch.py

@@ -2042,6 +2042,11 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
     def is_dllm(self):
         return self.dllm_config is not None
 
+    def grammar_needs_sync(self) -> bool:
+        """Whether grammar forces this batch onto the synchronous path, i.e. the
+        previous batch's result is resolved before this forward."""
+        return self.has_grammar and not self.spec_algorithm.supports_grammar_overlap()
+
     def prepare_encoder_info_extend(
         self, input_ids: List[array[int]], seq_lens: List[int]
     ):

python/sglang/srt/managers/scheduler.py

@@ -1672,14 +1672,12 @@ class Scheduler(
             and last_batch_is_extend
         )
 
-        # Spec algorithms that don't advance the grammar FSM inside verify() (see
+        # Sync so the FSM advance lands before the next batch's bitmask. Permanent
-        # supports_grammar_overlap) still need overlap forced off for grammar decode
+        # path for host-draft algorithms, not a pending migration.
-        # batches, so the FSM is advanced before the next batch's bitmask.
         need_grammar_sync = (
             batch
             and not batch.spec_algorithm.is_none()
-            and not batch.spec_algorithm.supports_grammar_overlap()
+            and batch.grammar_needs_sync()
-            and batch.has_grammar
             and batch.forward_mode.is_decode()
             and len(self.result_queue) > 0
         )

python/sglang/srt/speculative/ngram_worker.py

@@ -230,7 +230,7 @@ class NGRAMWorker(BaseSpecWorker):
         # spliced in from spec_info. Sync mode and grammar batches process
         # results before the next draft prep, so output_ids is already
         # complete and splicing would duplicate the tail.
-        use_prev_tokens = self.enable_overlap and not batch.has_grammar
+        use_prev_tokens = self.enable_overlap and not batch.grammar_needs_sync()
         i = 0
         for req in batch.reqs:
             prev_tokens = (
@@ -344,7 +344,7 @@ class NGRAMWorker(BaseSpecWorker):
         i, stride = 0, self.draft_token_num
         # Same splice condition as _prepare_draft_tokens: only overlap mode
         # has accepted tokens missing from req.output_ids.
-        use_prev_tokens = self.enable_overlap and not batch.has_grammar
+        use_prev_tokens = self.enable_overlap and not batch.grammar_needs_sync()
         for req in batch.reqs:
             # FIXME: Whether to insert 'extend' into the cache or not, after testing,
             # there is not much difference, so we will not insert it for now.

python/sglang/srt/speculative/spec_info.py

@@ -136,6 +136,8 @@ class SpeculativeAlgorithm(Enum):
     def supports_grammar_overlap(self) -> bool:
         # Whether the worker advances the grammar FSM inside verify() (via the
         # scheduler's grammar barrier), letting spec + grammar decode overlap.
+        # Needs a GPU draft phase to hide the grammar CPU work under: NGRAM drafts
+        # from a host corpus lookup, so it stays synchronous by design.
         # STANDALONE inherits the EAGLE V2 worker's verify path, barrier included.
         return self.is_eagle() or self.is_standalone()