Fix disagg overlap scheduler grammar sync for spec_v2 + guided decoding

python/sglang/srt/constrained/reasoner_grammar_backend.py

@@ -78,6 +78,7 @@ def __init__(
     def maybe_init_reasoning(self, reasoning: bool):
         if reasoning:
             self.tokens_in_think = 0
+            self.tokens_after_end = -1  # reset: cache hit may have stale GENERATION state
         else:
             self.tokens_in_think = -1
             self.tokens_after_end = 0

python/sglang/srt/disaggregation/decode.py

@@ -1682,6 +1682,15 @@ def event_loop_normal_disagg_decode(self: Scheduler):
             # Get the next batch to run
             batch = self.get_next_disagg_decode_batch_to_run()
             self.cur_batch = batch
+            disable_overlap_for_batch = self.is_disable_overlap_for_batch(batch)
+
+            # Match the normal overlap scheduler: for spec_v2 + grammar, commit
+            # accepted tokens to req.output_ids / req.grammar before launching the
+            # next decode batch. Otherwise disagg overlap can prepare the next MTP
+            # verify from advanced spec state while the grammar matcher is still
+            # one result behind.
+            if disable_overlap_for_batch and self.last_batch:
+                pop_and_process()
 
             # Launch the current batch
             if batch:
@@ -1699,6 +1708,10 @@ def event_loop_overlap_disagg_decode(self: Scheduler):
         self.result_queue = deque()
         self.last_batch: Optional[ScheduleBatch] = None
 
+        def pop_and_process():
+            tmp_batch, tmp_result = self.result_queue.popleft()
+            self.process_batch_result(tmp_batch, tmp_result)
+
         while True:
             # Receive requests
             recv_reqs = self.request_receiver.recv_requests()
@@ -1714,6 +1727,15 @@ def event_loop_overlap_disagg_decode(self: Scheduler):
             # Get the next batch to run
             batch = self.get_next_disagg_decode_batch_to_run()
             self.cur_batch = batch
+            disable_overlap_for_batch = self.is_disable_overlap_for_batch(batch)
+
+            # Match the normal overlap scheduler: for spec_v2 + grammar, commit
+            # accepted tokens to req.output_ids / req.grammar before launching the
+            # next decode batch. Otherwise disagg overlap can prepare the next MTP
+            # verify from advanced spec state while the grammar matcher is still
+            # one result behind.
+            if disable_overlap_for_batch and self.last_batch:
+                pop_and_process()
 
             # Launch the current batch
             if batch:
@@ -1724,8 +1746,8 @@ def event_loop_overlap_disagg_decode(self: Scheduler):
 
             # Process the last batch
             if self.last_batch:
-                tmp_batch, tmp_result = self.result_queue.popleft()
-                self.process_batch_result(tmp_batch, tmp_result)
+                if not disable_overlap_for_batch:
+                    pop_and_process()
             elif batch is None:
                 self.on_idle()