sd: sync to master-387-e4c50f1

[wbruna]

Almost no code conflicts on this one.

I had to add new wrappers on model.cpp to load the new Mistral vocabulary and merges. I don't really have the hardware to test Flux.2, though, so I didn't try to include the matching .embd files.

Qwen loading is broken, probably due to the qwen2vl -> llm replacement. I'll see if I can fix it.

[wbruna]

I didn't hit other obvious bugs, so I think this is ready for review.

[LostRuins]

Qwen loading is broken, probably due to the qwen2vl -> llm replacement

wait what happened? didn't notice they touched qwen

[wbruna]

wait what happened? didn't notice they touched qwen

qwen2vl was generalized to llm all around. It didn't cause a build failure on Koboldcpp, but the loaded qwen2vl tensor prefix changed from "text_encoder.qwen2vl." to "text_encoder.llm." (look for "text_encoders." on leejet/stable-diffusion.cpp@52b67c5), so our clip -> qwen replacement didn't work anymore.

[wbruna]

I'll work on it in a few hours.

By the way, leejet just merged the Z-Image-Turbo support. Shall I include it here, or on a follow-up PR? (the only problem with a separate PR would be the fix for Flux.2 VAE tiling, but I can include it first).

[LostRuins]

We can combine them I think. At most we can do the same approach we did previously where we eventually squash it into two commits and I do a merge

[wbruna]