Endless repetition in output

[andypotato]

Description

Using the 1.6b version served in a docker container based on vllm/vllm-openai:v0.17.1.

It works fine for most files. However sometimes the audio input will cause the model to spiral out of control and repeat the same token over and over again. Here is an example:

Are there going to be exercise part of things that they can exercise? I think. I think. I think. I think. I think. I think. I think. I think. I think. I think. I think. I think. I think. I think. I think. I think.
[...]

In this example it repeats "I think" for about 2,000 times before it continues the transcription.

Is there any way to avoid this issue?

Reproduction

Reproducible with the same audio file. About 1 in 10 files will show this behavior. Original language is Chinese.

Logs

Environment Information

Docker (vllm/vllm-openai:v0.17.1)

Known Issue

• The issue hasn't been already addressed in Documentation, Issues, and Discussions.