Fix spec info's filter when reqs are finished right after prefill (#14742)

python/sglang/srt/managers/schedule_batch.py

@@ -1796,6 +1796,8 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         self,
         chunked_req_to_exclude: Optional[Union[Req, List[Req]]] = None,
         keep_indices: Optional[List[int]] = None,
+        # FIXME(lsyin): deprecate this API after spec v1 is deprecated
+        v1_spec_info_filtered: Optional[bool] = False,
     ):
         # FIXME(lsyin): used here to get the correct seq_lens
         # The batch has been launched but we need it verified to get correct next batch info
@@ -1852,11 +1854,12 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         self.has_grammar = any(req.grammar for req in self.reqs)
 
         self.sampling_info.filter_batch(keep_indices, keep_indices_device)
+        # NOTE: spec_info filtered before batch filtering only happens in:
+        # - Spec v1's verify phase
+        # - Only for decode batch (running_batch)
+        has_been_filtered = v1_spec_info_filtered and not self.is_v2_eagle
+
         if self.spec_info:
-            if chunked_req_to_exclude is not None and len(chunked_req_to_exclude) > 0:
-                has_been_filtered = False
-            else:
-                has_been_filtered = True
             self.spec_info.filter_batch(
                 new_indices=keep_indices_device,
                 has_been_filtered=has_been_filtered,

python/sglang/srt/managers/scheduler.py

@@ -1937,7 +1937,7 @@ class Scheduler(
             and not (new_batch.return_logprob or self.running_batch.return_logprob)
         ):
             # TODO (lianmin): support return_logprob + mixed chunked prefill
-            self.running_batch.filter_batch()
+            self.running_batch.filter_batch(v1_spec_info_filtered=True)
             if not self.running_batch.is_empty():
                 self.running_batch.prepare_for_decode()
                 new_batch.mix_with_running(self.running_batch)
@@ -1954,7 +1954,7 @@ class Scheduler(
         """Update the current running decoding batch."""
         initial_bs = batch.batch_size()
 
-        batch.filter_batch()
+        batch.filter_batch(v1_spec_info_filtered=True)
         if batch.is_empty():
             batch.batch_is_full = False
             return batch
@@ -2509,7 +2509,7 @@ class Scheduler(
         self.cur_batch = None
 
         if recv_req.mode == "retract":
-            self.running_batch.filter_batch()
+            self.running_batch.filter_batch(v1_spec_info_filtered=True)
             if len(self.running_batch.reqs) != 0:
                 retracted_reqs = self.running_batch.retract_all(self.server_args)
                 for req in retracted_reqs: