diff --git a/vllm/core/scheduler.py b/vllm/core/scheduler.py
index c57e6cd71640..5cdb490e305f 100644
--- a/vllm/core/scheduler.py
+++ b/vllm/core/scheduler.py
@@ -1202,10 +1202,11 @@ def _can_append_slots(self, seq_group: SequenceGroup,
             seq_group=seq_group, num_lookahead_slots=num_lookahead_slots)
 
     def _allow_async_output_proc(self, seq_group: SequenceGroup) -> bool:
-        # TODO: does it work with parallel sampling?
-        no_beam_search = seq_group.sampling_params is None or (
+        # async_output_proc is allowed only when we have a single sequence
+        # in the sequence group
+        no_single_seq = seq_group.sampling_params is None or (
             seq_group.sampling_params.best_of == 1)
-        return no_beam_search
+        return no_single_seq
 
     def schedule(
             self