[Feature]: supported yarn rope-scaling

[15626471095]

🚀 The feature, motivation and pitch

--rope-scaling '{"rope_type":"yarn","factor":4.0,"original_max_position_embeddings":32768}' --max-model-len 131072

Extend qwen3 context length from 32768 to 131072

Alternatives

No response

Additional context

No response

[zhanglzu]

If this is a feature request, then vLLM already supports it, at least in vLLM Ascend v0.7.3.post1.

[paulyu12]

I tested the "yarn rope scaling" feature with factors 2.0 and 4.0 on vllm-ascend:main + Qwen3-14B, and found that both ran well. How to reproduce your issue ?

[Yikun]

Yes, it's ready supported, see also here: https://qwen.readthedocs.io/en/latest/deployment/vllm.html#context-length