feat: support new chroma radiance "x0_x32_proto"

[stduhpf]

https://huggingface.co/lodestones/Chroma1-Radiance/tree/main

I also made sure to keep compatibility with the older "x0_experimental_patch_32" model (this one was using 2x downscaling before 16x16 patches instead of actual 32x32 patches like the new one).

[leejet]

Thank you for your contribution.

[Green-Sky]

@stduhpf how did you get the models to convert?
I think ggml really does not like zero sized tensors. They are also technically unsupported.

---

I wanted to check out the latest. It is apparently now training on 1024x1024.

[stduhpf]

@Green-Sky b9da97c#diff-6ccc4ebe687dcee80558657369de532fabd66d1804275b79f7a030af880d5760R1736-R1737

[Green-Sky]

Odd, I have that. I tried various states of this repo now. I always get some variant of

Thread 2 "sd-cli" received signal SIGFPE, Arithmetic exception.

When I try to convert.

edit: with a bogus source code location. (yes without optimizations too)

[stduhpf]

Weird, it still seems to work fine on my machine. Maybe that's a windows vs posix thing.

What if you set tensor->ne to be all 1 in model.cpp line 1753 onwards?

Edit: tried it on WSL, I can confirm it doesn't work there.

[Green-Sky]

asan/ubsan tells me there is a division by zero here:

stable-diffusion.cpp/src/model.cpp

Line 1542 in 636d3cb

(int)tensor_storage.nelements() / (int)tensor_storage.ne[0],

[stduhpf]

Well then i don't think there is an easy fix or workaround we could use on the stable-diffusion.cpp side to make it work on Linux like it does on windows. The best way to handle it would probably to add proper support for zero-sized tensors in ggml directly.

[Green-Sky]

I would feel more comfortable with metadata. Or if we really need it single scalars. ggml/gguf walked away from zero sized tensors.

---

Oh and I got it to work anyway by adding a ne[0] > 0 check before that conversion line.

[stduhpf]

Oh, you're right, this works. For some reason I assumed the division by zero had to happen in ggml and didn't even check your comment properly.

[Green-Sky]

I am a little confused about flash attention. It seems that it does not reduce the memory consumption at all. It does how ever slightly speed things up by ~10%. Kinda odd.

Also cool to see the image compose with --preview vae, seems like the last few steps end up destroying the image somewhat.