Fallback to orignal model

[YuchaoXu]

when i use Evalplus to evaluate the qwen3-flashhead model,it can not correctly applied, and i think it fallback to qwen3.but when i set "export VLLM_ENABLE_V1_MULTIPROCESSING=0",it's successfully applied

[JonnaMat]

Thanks for reporting this, and good find on the workaround!

This is a known limitation with how the patches propagate in vLLM's V1 multiprocessing mode: the worker processes don't pick up the modifications when they spawn in separate processes.

We've since moved FlashHead to a standalone vLLM plugin that registers via the official vllm.general_plugins entry point, which should handle multi-process discovery correctly:

  pip install flash-head

This replaces embedl-models, which will be deprecated shortly in favor of flash-head. FlashHead now activates automatically in every vLLM process (including workers). Could you try with the new plugin and let us know if the V1 multiprocessing issue persists?