[fix]修复mtp的1处笔误

vllm/v1/worker/gpu_model_runner.py

@@ -1796,13 +1796,14 @@ class GPUModelRunnerBase(LoRAModelRunnerMixin):
                 decoding=spec_decode_metadata is not None
             )
 
-            spec_token_ids = draft_token_ids.tolist()
-
             if not envs.VLLM_REJECT_SAMPLE_OPT:
                 draft_token_ids = draft_result
             else:
-                draft_req_ids = list(scheduler_output.num_scheduled_tokens.keys())
                 draft_token_ids, draft_probs = draft_result
+            spec_token_ids = draft_token_ids.tolist()
+
+            if envs.VLLM_REJECT_SAMPLE_OPT:
+                draft_req_ids = list(scheduler_output.num_scheduled_tokens.keys())
                 if self.draft_probs is None:
                     self.draft_probs = DraftProbs(
                         draft_probs, draft_req_ids)