[cbk-res]

Description

This PR attempts to modernise support for stable-diffusion.cpp within llamafile through diffusionfile. I have updated the submodule in use to the latest branch (baf7eda1), removed the now redundant and obsolete patches which were needed by the old version, but seemingly are resolved at present. The new implementation patches SDcpp minimally, following the approach used for whispercpp, and in doing so, removes substantial amounts of code, and therefore maintenance burden. It is working well for me locally as you can see below (although only CPU tested so far), testing with SDXS Q8 gguf, and everything seems to work as expected on my system.

./diffusionfile -m sdxs-512-tinySDdistilled_Q8_0.gguf -p "a happy donkey" --steps  8

PR Type

• 🆕 New Feature
• 🐛 Bug Fix

Relevant issues

#827
#229

Checklist

• I understand the code I am submitting.
• I have run this code locally and verified the change.
• New and existing tests pass locally, or I have explained why tests were not run.
• Documentation was updated where necessary.
• If I changed code in llama.cpp/, whisper.cpp/, or stable-diffusion.cpp/, I also updated the matching *.patches/ files.
• I have read and followed the contribution guidelines.
• AI Usage:
  • No AI was used.
  • AI was used in an assistive capacity.
  • This PR includes substantial AI-generated content.

AI Usage Information

• AI Model used: Qwen 3.6 Plus
• AI Developer Tool used: OpenCode
• Any other info you'd like to share: This code has been implemented by my agent under strict instructions to remain firmly within the conventions used in stock llamafile, and as such, the modernisation of diffusionfile follows closely that of the implementation used for whisperfile. I have personally reviewed this code, and had a second human review also. This code has been tested extensively locally, as well as using using the automated tests (integrations pytests - although the nothink shortened output one times out on my cpu, and URI test) to verify that stock functionality is unaffected, and the SDcpp implementation with SDXS, to ensure that all is in working order. However, I do not have a GPU, so testing of this functionality is unverified.

When answering reviewer questions, please respond yourself rather than pasting reviewer comments into an AI system and posting the reply back unchanged.

• I am an AI Agent filling out this form (check box if true)

[jart]

These should be defined as:

long long llround(double x) {
  return round(x);
}
long long llroundf(float x) {
  return roundf(x);
}
long long llroundl(long double x) {
  return roundl(x);
}

I'm adding them to cosmo libc next release.

[cbk-res]

Thanks for the review, this should be corrected now :)
I built with the changes and it seems to work as expected.

[aittalam]

Hi @cbk-res ! Many thanks for your contribution 🙏

Given this adds back a feature we had lost, I think it's great even if it does support CPU alone and we can merge it as-is (similarly to what we did with whisperfile) then work to bring GPU support back into it.

Code runs well on my system too, from setup to building and testing it

[cbk-res]

No problem at all @aittalam, glad it is working well. I hope to contribute more in the future :)