A bug about the web ui and the llm model.

[idiotTest]

Describe the bug

I found a bug is that the gpu memory shows the xinf and the llm model are running,but i can't see the model in web ui.Also,I can't use the llm model.

To Reproduce

At begining,I run qwen1.5-14b-chat in a linux with xinf,llvm.And i do a test,I want to do a stress test to test some indicators.At first time,it seems ok.But when i run again,I found the bug.
Also,the customer model registered by user is also lost

###Some guesses
The gpu memory usage exceeds and cause the error?