fix: 解决列表访问越界

vllm/zero_overhead/v1/core.py

@@ -72,6 +72,8 @@ def zero_overhead_update_from_output(scheduler:Scheduler,
                 continue
             request = scheduler.requests[req_id]
             generated_token_ids = model_runner_output.fix_sampled_token_ids[req_idx]
+            if not generated_token_ids:
+                continue
             if req_id not in requsets_valid_token_len:
                 requsets_valid_token_len[req_id] = 0
             valid_output_len = requsets_valid_token_len[req_id]

vllm/zero_overhead/v1/gpu_model_runner.py

@@ -705,6 +705,7 @@ class V1ZeroModelRunner(GPUModelRunner):
                 self.fix_sampled_token_ids = self.last_sampler_host_tokens.tolist()
                 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
                     req_id = self.fix_req_ids[req_idx]
                     if req_id in self.input_batch.req_ids: