Why kohya_ss inference speed faster than diffusers or Onetrainer?

I trained a lora with kohya_ss and Onetrainer, I'm confused that why the lora with kohya_ss is much faster than using diffusers?
If the network is different from others, why diffusers state dict can convert into kohya_ss?

Compare inference behavior for equivalent LoRAs trained with kohya_ss, Onetrainer, and diffusers, starting with the state-dict conversion path mentioned in the issue. Determine whether the reported speed difference comes from network structure or conversion, and document a reproducible explanation; no files or tests are named.

Written by the indexing model from the issue text.