[Fix] Fix best_of behavior when n=1 (#3298)

vllm/outputs.py

@@ -87,12 +87,12 @@ class RequestOutput:
 
     @classmethod
     def from_seq_group(cls, seq_group: SequenceGroup) -> "RequestOutput":
-        # Get the top-n sequences.
-        n = seq_group.sampling_params.n
         seqs = seq_group.get_seqs()
-        if n == 1:
+        if len(seqs) == 1:
             top_n_seqs = seqs
         else:
+            # Get the top-n sequences.
+            n = seq_group.sampling_params.n
             if seq_group.sampling_params.use_beam_search:
                 sorting_key = lambda seq: seq.get_beam_search_score(
                     seq_group.sampling_params.length_penalty)