How to use Smoothquant and FP16 or Weight Only in a LLM on the same time?

[micyan01]

We know that smoothquant has little precision loss.
Could we choose some layers to quantitize using Smoothquant and others to keep float16. Is that work, and how to config.