[Performance]: yarn degrades the performance of qwen3

[yanan1116]

Proposal to improve performance

vllm version == 0.8.5.post1

without yarn

vllm serve Qwen/Qwen3-32B   \
 --trust-remote-code --gpu_memory_utilization 0.95 --tensor-parallel-size 2 \
--quantization bitsandbytes --load_format bitsandbytes --enforce_eager \
--max-model-len 32768

with yarn

vllm serve Qwen/Qwen3-32B   \
--trust-remote-code --gpu_memory_utilization 0.95 --tensor-parallel-size 2 \
--quantization bitsandbytes --load_format bitsandbytes --enforce_eager \
--rope-scaling '{"rope_type":"yarn","factor":4.0,"original_max_position_embeddings":32768}' \
--max-model-len 131072

I have some tests on my end for its agentic capabilities based on qwen3 and I have some solid findings that enabling yarn to extend window context does degrade the performace, with around 15-20% performance drop.

do u also encounter the same findings ? any suggestion about this drop ?

Report of performance regression

No response

Misc discussion on performance

No response

Your current environment (if you think it is necessary)

The output of `python collect_env.py`

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.

[yanan1116]

seem that it is because of static YaRN ?
https://qwen.readthedocs.io/en/latest/deployment/vllm.html