@leejet Due to the API refactor, I think I don't need to add additional package code (I mean this project https://github.com/seasonjs/stable-diffusion.cpp-build). Are you willing to let github ci support building dynamic libraries? If you are willing, I am willing to modify ci and submit a PR.
Below is my planned dynamic library compatibility table:

@Cyberhan123 is there a point in not using avx2 for linux/mac amd64 ?

@Cyberhan123 is there a point in not using avx2 for linux/mac amd64 ?

Only supports avx2

@leejet would be cool if we could bake in a custom vae, like the f16 fix for xl.

Only supports avx2

I use avx2 and CUDA with Linux.

Are you willing to let github ci support building dynamic libraries?

@Cyberhan123 I'd appreciate someone helping to set up the corresponding CI. I haven't had much free time lately to work on this.

@leejet would be cool if we could bake in a custom vae, like the f16 fix for xl.

I'm not quite following what you mean. It's been possible to specify custom vae for a while now.

@leejet would be cool if we could bake in a custom vae, like the f16 fix for xl.

I'm not quite following what you mean. It's been possible to specify custom vae for a while now.

but to have them merged/replaced into the same gguf

@leejet would be cool if we could bake in a custom vae, like the f16 fix for xl.

I'm not quite following what you mean. It's been possible to specify custom vae for a while now.

but to have them merged/replaced into the same gguf

It has been supported now.

.\bin\Release\sd.exe -M convert -m ..\..\stable-diffusion-webui\models\Stable-diffusion\sd_xl_base_1.0.safetensors --vae ..\..\stable-diffusion-webui\models\VAE\sdxl_vae-fp16-fix.safetensors -o ..\models\sd_xl_base_1.0_with_vae_fix.gguf -v

.\bin\Release\sd.exe -m ..\models\sd_xl_base_1.0_with_vae_fix.gguf -p "a lovely cat" -v   -H 1024 -W 1024

I'd appreciate someone helping to set up the corresponding CI. I haven't had much free time lately to work on this.

Okay, after ggml is merged and the PR of rocm is completed, I will implement it.