[Bug]: Get meaningless output when run long context inference of Qwen2.5 model with vllm>=0.6.3

[piamo]

Your current environment

The output of `python collect_env.py`

Your output of `python collect_env.py` here

Model Input Dumps

models: Qwen2.5-Coder-7B-Instcut, Qwen2.5-7B-Instruct
vllm: 0.6.3
input token: >8000 tokens

🐛 Describe the bug

I have tested vllm 0.6.0~0.6.2, 0.5.5, all old versions are just ok.

So this bug was introduced since 0.6.3

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[CHNtentes]

same here, used qwen2.5-72b-instruct-awq and 10000 tokens input, the output is garbage

[CHNtentes]

downgraded to vllm 0.6.2 and it's much better