[Bug]:vllm从0.7.0开始版本部署Qwen2_vl服务存在内存(不是GPU显存)泄漏问题

[Syx93-c]

Your current environment

vllm从0.7.0开始版本部署Qwen2_vl服务存在内存(不是GPU显存)泄漏问题

🐛 Describe the bug

使用0.7.0版本的vllm部署Qwen2_vl模型服务时，对服务进行请求后，服务相关进程内存不会释放，最终打爆服务器内存，导致服务停止，测试发现0.6.6版本无此问题，0.7.0及以上版本均有此问题。

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.

[Syx93-c]

基础配置cuda12.4 python 3.10

torch 2.5.1
torchaudio 2.5.1
torchvision 0.20.1
tornado 6.4.2
tqdm 4.67.1
traitlets 5.14.3
transformers 4.49.0
triton 3.1.0
typer 0.15.2
types-python-dateutil 2.9.0.20241206
typing_extensions 4.12.2
tzdata 2025.1
uri-template 1.3.0
urllib3 2.3.0
uvicorn 0.34.0
uvloop 0.21.0
virtualenv 20.29.2
vllm 0.7.0
vllm-flash-attn 2.6.1
watchfiles 1.0.4
wcwidth 0.2.13
webcolors 24.11.1
webencodings 0.5.1
websocket-client 1.8.0
websockets 15.0
wheel 0.43.0
wrapt 1.17.2
xformers 0.0.28.post3
xgrammar 0.1.16
xxhash 3.5.0
yarl 1.18.3
zipp 3.21.0

[DarkLight1337]

This issue should be fixed from v0.8.2 onwards, can you try it?

[Syx93-c]

This issue should be fixed from v0.8.2 onwards, can you try it?

I tried it but it didn't work. v0.8.2 still have a memory leak.

[DarkLight1337]

Do you get similar issues with any other models?

[Syx93-c]

Do you get similar issues with any other models?

At present, I have mainly tried the Qwen2_VL and Qwen2.5VL series, they have the problem,.
and have not tried other

[Syx93-c]

Do you get similar issues with any other models?

I can quickly try another model, wait a moment

[DarkLight1337]

There is increased memory usage in recent versions due to the addition of preprocessing cache. However the memory usage should stabilize after a while rather than increasing indefinitely. If you get OOM issues, these options may be able to help:

• Set VLLM_MM_INPUT_CACHE_GIB environment to a smaller value (the default is 4 which means 4 GiB)
• Disable the cache completely by --disable-mm-preprocessor-cache

[Syx93-c]

Do you get similar issues with any other models?

I can quickly try another model, wait a moment

MiniCPM have the same problem

[Syx93-c]

There is increased memory usage in recent versions due to the addition of preprocessing cache. However the memory usage should stabilize after a while rather than increasing indefinitely. If you get OOM issues, these options may be able to help:

• Set VLLM_MM_INPUT_CACHE_GIB environment to a smaller value (the default is 4 which means 4 GiB)
• Disable the cache completely by --disable-mm-preprocessor-cache

ok I will try it,thanks a lot!

[Syx93-c]

There is increased memory usage in recent versions due to the addition of preprocessing cache. However the memory usage should stabilize after a while rather than increasing indefinitely. If you get OOM issues, these options may be able to help:

• Set VLLM_MM_INPUT_CACHE_GIB environment to a smaller value (the default is 4 which means 4 GiB)
• Disable the cache completely by --disable-mm-preprocessor-cache

ok I will try it,thanks a lot!

I used --disable-mm-preprocessor-cache it works! thank you very much!

[wangyeqiang]

遇到同样问题，调整 --disable-mm-preprocessor-cache 仍然无效

[DarkLight1337]

Which version of vLLM are you on? You should upgrade to v0.8.2.