Merge branch 'v0.9.2-dev' of http://10.16.6.30/dcutoolkit/deeplearing/vllm into v0.9.2-dev

vllm/zero_overhead/v1/core.py

@@ -177,11 +177,11 @@ def zero_overhead_update_from_output(scheduler:Scheduler,
     # loop can be a performance bottleneck. We should do our best to avoid
     # expensive operations inside the loop.
     for request in scheduler.running:
+        req_id = request.request_id
         if request.is_finished():
             if req_id in requsets_valid_token_len:
                 requsets_valid_token_len.pop(req_id)
             continue
-        req_id = request.request_id
         num_tokens_scheduled = num_scheduled_tokens.get(req_id, 0)
         if num_tokens_scheduled == 0:
             # The request was not scheduled in this step.