[Bug] 使用max-num-worker使得ssh断连

[timturing]

Prerequisite

• [x] I have searched Issues and Discussions but cannot get the expected help.
• [x] The bug has not been fixed in the latest version.

Type

I'm evaluating with the officially supported tasks/models/datasets.

Environment

我需要并行评测的我的模型，因为评估单个example的时间太长，所以我参考了 #1755 中的代码，使用CUDA_VISIBLE_DEVICES=6,7 opencompass --models vllm_qwen2_5_0_5b_instruct --datasets triviaqa_gen -a vllm --max-num-worker 2这种方式进行并行。然而每次当我启动此代码时，这将导致我的ssh直接断连，并返回报错shell request failed on channel 0，感觉该问题与opencompass的code有关。

Reproduces the problem - code/configuration sample

CUDA_VISIBLE_DEVICES=6,7 opencompass --models vllm_qwen2_5_0_5b_instruct --datasets triviaqa_gen -a vllm --max-num-worker 2

Reproduces the problem - command or script

CUDA_VISIBLE_DEVICES=6,7 opencompass --models vllm_qwen2_5_0_5b_instruct --datasets triviaqa_gen -a vllm --max-num-worker 2

Reproduces the problem - error message

shell request failed on channel 0

Other information

[tonysy]

Can you launch the evaluation with --debug?

[timturing]

Yes, when running with --debug option, the ssh connection is fine. However, it will only use one GPU instead of the number I set in CUDA_VISIBLE_DEVICES and max-num-worker.

[tonysy]

Actually, I have no idea about this bug. Can you use tmux and then try to remove --debug?