Offloading feature

[eeditor1055]

Few days ago I saw a post on reddit, where people offload LLM GGUF not full layer by full layer to GPU, but certain parts of layers to GPU.
They claim that It boosts speed of generation.

Can this be done with converted SD (SD1.5, SDXL, FLUX, etc..) GGUFs?

Here's a link to the post:
https://www.reddit.com/r/LocalLLaMA/comments/1ki7tg7/dont_offload_gguf_layers_offload_tensors_200_gen/