Raw default output for many model :

Result with this PR :

Sorry for the spam : I had two versions coexisting in my Git history, so I reverted to the simple version that works on its own, without the remark plugin.

This patch renders correctly mathematical expressions on my end.

Just a minor note - probably the "Show raw LLM output" option should not be located in the "Reasoning" tab. And the name should be shortened to just "Show raw LLM output".

I realized it and told Alek on discord:D

Because we should group all the debug options into a dedicated developer section, since we’re going to add toolcall debugging soon.
And also "Custom JSON parameters to send to the API. Must be valid JSON format." -> Debug or Developer...
And eventually, the “Advanced” section could be renamed to “Sampling.”
And current "Sampler" inside "Sampling"
What do you think?

Settings layout : #16607

Hi @ServeurpersoCom , in #16508 there are test-samples in https://github.com/ggml-org/llama.cpp/pull/16508/files#diff-22e3b93a74a89292e31a9c73b38e1e011c26b28f0715369f30f5a8d1090ea750 .

Hi @ServeurpersoCom , in #16508 there are test-samples in https://github.com/ggml-org/llama.cpp/pull/16508/files#diff-22e3b93a74a89292e31a9c73b38e1e011c26b28f0715369f30f5a8d1090ea750 .

I’ll take a look at your test samples: that seems like a really good idea.

Take a look at my proposal: it’s much simpler now, just one function. Implementing it as a remark plugin was making things more complex.

The remark MathJax plugin rewrites the Markdown AST too aggressively, replacing whole subtrees and risking broken output whenever nested markup or edge cases appear. A simple regex-based pre-normalization that converts math delimiters before parsing avoids all that complexity, it lets the standard remark-math plugin do the job safely. Fewer moving parts mean less maintenance, fewer bugs, and the same rendering quality without the AST surgery.