Qwen3.8 27B with embeddings

As you know, Qwen3.8 27B has built-in support for embeddings. It gives pretty decent results in llama.cpp with the following parameters:

--embedding
--pooling mean

The problem is that enabling embeddings cuts the token generation speed in half compared to running the model normally.

Has anyone experimented with this? Is there a way to keep normal generation performance while having embeddings enabled? Or can llama.cpp switch between different tasks without having to unload and reload the model every time?