[fix]解决同时开启chunked-prefill和并行解码出现的指标统计错误问题

vllm/v1/core/sched/scheduler.py

@@ -1225,7 +1225,7 @@ class Scheduler(SchedulerInterface):
                 spec_decoding_stats = self.make_spec_decoding_stats(
                     spec_decoding_stats,
                     num_draft_tokens=len(scheduled_spec_token_ids),
-                    num_accepted_tokens=len(generated_token_ids) - 1)
+                    num_accepted_tokens=max(0, len(generated_token_ids) - 1))
 
             # NOTE(woosuk): This has to be executed after updating
             # `request.num_computed_tokens`.