fix: chunk-prefill生成的token被错误返回

vllm/zero_overhead/v1/gpu_model_runner.py

@@ -98,6 +98,7 @@ class V1ZeroModelRunner(GPUModelRunner):
                 )
             for req_idx, start_idx, end_idx in self.token_ids_cpu_fix_record:
                 if start_idx == -1:
+                    self.fix_sampled_token_ids[req_idx].clear()
                     continue
                 num_accepted_tokens = len(self.fix_sampled_token_ids[req_idx])
                 req_id = self.fix_req_ids[req_idx]