请问作者在实验过程中出现过一直运行不停止的情况吗。执行了20多轮还不停

[expatriate12138]

[zhangshaolei1998]

这个是一直出现么？我们在测试过程中很少遇到这种情况。

[expatriate12138]

是的，后端llm也一直在输出同样的内容，和vllm部署的参数有关吗

…

------------------ 原始邮件 ------------------ 发件人: ZhangShaolei ***@***.***> 发送时间: 2025年11月13日 12:25 收件人: ruc-datalab/DeepAnalyze ***@***.***> 抄送: expatriate12138 ***@***.***>, Author ***@***.***> 主题: Re: [ruc-datalab/DeepAnalyze] 请问作者在实验过程中出现过一直运行不停止的情况吗。执行了20多轮还不停 (Issue #17)

[zhangshaolei1998]

是的，后端llm也一直在输出同样的内容，和vllm部署的参数有关吗
…

有可能，vllm要大于0.8.5。另外，您使的是全量版本还是量化后的呢？

[wynshiter]

我的情况也是运行了一百多轮次，一直不结束

提示词是介绍背景+

Please output:
Key driver features and their importance scores.
Visualizations (charts/tables).
Actionable recommendations.
数据是一份原始数据和一份说明

下面是我vllm 的参数

(deepanalyze) season@WYANING:~/DeepAnalyze-main/demo$ vllm serve /home/season/.cache/modelscope/hub/models/RUC-DataLab/DeepAnalyze-8B --dtype bfloat16 --enforce-eager --served-model-name DeepAnalyze-8B --gpu-memory-utilization 0.5 --max-model-len 4096 --max-num-seqs 1 --max-num-batched-tokens 256 --kv-cache-dtype fp8

下面是一部分vllm 日志

(APIServer pid=904) INFO: 127.0.0.1:49316 - "POST /v1/chat/completions HTTP/1.1" 200 OK
(APIServer pid=904) INFO 11-14 22:13:13 [loggers.py:127] Engine 000: Avg prompt throughput: 354.5 tokens/s, Avg generation throughput: 31.6 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.1%, Prefix cache hit rate: 99.6%
(APIServer pid=904) INFO 11-14 22:13:23 [loggers.py:127] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 23.5 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 99.6%
(APIServer pid=904) INFO 11-14 22:13:33 [loggers.py:127] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 0.0 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 99.6%
(APIServer pid=904) WARNING 11-14 22:18:01 [protocol.py:93] The following fields were present in the request but ignored: {'max_new_tokens'}
(APIServer pid=904) INFO: 127.0.0.1:47668 - "POST /v1/chat/completions HTTP/1.1" 200 OK
(APIServer pid=904) INFO 11-14 22:18:03 [loggers.py:127] Engine 000: Avg prompt throughput: 354.5 tokens/s, Avg generation throughput: 9.2 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.9%, Prefix cache hit rate: 99.6%
(APIServer pid=904) INFO 11-14 22:18:13 [loggers.py:127] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 44.6 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.3%, Prefix cache hit rate: 99.6%
^C(APIServer pid=904) INFO 11-14 22:18:19 [launcher.py:99] Shutting down FastAPI HTTP server.
[rank0]:[W1114 22:18:19.967849069 ProcessGroupNCCL.cpp:1538] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
(APIServer pid=904) INFO: Shutting down
(APIServer pid=904) INFO: Waiting for application shutdown.
(APIServer pid=904) INFO: Application shutdown complete.

后台日志没有什么异常 请问一下可能的原因是什么呢 @zhangshaolei1998

[lordk911]

感觉和限制的上下文的长度有关，demo里的代码 "max_new_tokens": 32768,
如果总的上下文比较少，可能输出的标签不完整，就影响了demo程序对模型输出的判断。

[Soh-AI-B]

may be the problem comes from that the max_new_tokens is ignored by the vllm server since when runing the model it shows this
'WARNING 12-08 10:14:12 [protocol.py:112] The following fields were present in the request but ignored: {'max_new_tokens'}'