[BUG?] Output in web (kobold lite) became unformatted, then in red font midway then again normal

[alex-ie]

Describe the Issue
All setting were on defaults (except network interface port and number of CPU cores maybe - I used CPU here). My prompt was asking to calculate number of models parameters from copy-paste of kcpp output and then I repeatedly clicked 'Generate More'.
Output in web (kobold lite) run normal then in the middle of output (still white font):

$$
\text{Attention params} = 
3 \times g \time h_{\text{per group}} \time s^2 + d^2

Then for next and many lines after font is red (with couple of lines in between as white). After many lines started to be normal formatting again, last red line is:

#### Attention P...

I've checked output in linux terminal and the change from white to red is about 5 lines of text at end of one run of 'Generate More'.
part corresponding to change from white to red is (plus 4 lines more after Because):

> $$
\text{Attention params} = 
3 \times g \time h_{\text{per group}} \times s^2 + d^2
$$

Because:

I'm still conversing with the model. What additional info do you need? Is it known issue?

Additional Information:
Model: Qwen3-4B-Instruct
kcpp v1.110 linux64
Firefox
Linux Mint

TL;DR
I'm running in parallel Qwen 3.5 (with same kcpp file) answering same question (same have to click 'Gen More' again and again), making calculations and formatting of formulas etc is normal. At the change from white to yellow font (previous and current generation run) I saw
white:
$d_{model}
yellow:
/ 8 =320$
then formatting was "correct" again. After generation completed output in web corrected to 'proper' formatting along with font change to white. I guess it's how it's supposed to work.

[LostRuins]

Seems like a latex parsing issue. Do you have the input that causes it to break? I can take a look

[alex-ie]

Red-text-font-story.json

@LostRuins , attached is the saved story. I've also loaded it in v 1.111.2 - also seen ill-formatted red text in the middle.

[LostRuins]

Looks like its due to this block being misaligned for some reason

Note the blockquote > prefix

I fixed two of the cases, the final one that's fully blockquoted i can't fix properly.

[alex-ie]

@LostRuins

I fixed two of the cases, the final one that's fully blockquoted i can't fix properly.

Do you mean in the story there are another 2 blocks similar to the one you quoted?

i can't fix properly.

Do you mean you have no idea how now and the issue probably gonna linger for some time, correct?

[LostRuins]

Yes, in the original example there were 2 types of cases. The one that starts with a single blockquote i fixed, but i haven't found a robust regex that handles the multiline blockquote one without messing other stuff up.

[alex-ie]

haven't found a robust regex that handles the multiline blockquote one without messing other stuff up.

@LostRuins , not promising anything but I had some experience with regex and could try to help if you give a link and a bit of comment where to look. Do you use perl compatible type of regex or which?

[LostRuins]

Great. Check out the stashLatex and replaceLatex functions, around this line:

	text = text.replace(/(^> \$\$\n([\s\S]*?)\n\$\$)/gm, (match) => {
				const key = `%%LATEXBLK${counter++}%%`;
				latexBlocks.push({ key, value: match.slice(2) });
				return key;
			});

The raw LLM outputs something like this:

Actually, standard parameter count for **GQA** in papers is:

> Attention layer parameters per layer:
> $$
> 3 \times g \times h_{\text{per group}} \times s^2 + d^2
> $$

Where:
- `s` = head size
- `g` = number of groups
- `h_per_group` = number of heads per group

The idea is this: The LaTeX renderer does not recognize BlockQuotes >. So if they get sent into the LaTeX parser it will fail.
We need to detect this specific pattern of blockquotes (only containing pure latex) and then extract out those blockquotes in the stashLatex function. Later when the renderer sees the LaTeX it will be pure LaTeX and will be correctly rendered.

stashLatex is important because it transforms latex into escaped blocks so it doesnt corrupt other markdown when those are rendered. Ideally you'd add on to my above code to handle this special case, and stash the latex without the corrupting blockquotes.

$$
3 \times g \times h_{\text{per group}} \times s^2 + d^2
$$

Note that blockquotes can contain other things (including non-latex) so the regex must be very specific.

[alex-ie]

@LostRuins , I have looked at your code. I have little experience with programming in html.

I see commit 64d9531bc2e2c7dccfa54003fa3447a05d8c0247 (https://github.com/LostRuins/lite.koboldai.net) added:

text.replace(/(^> \$\$\n([\s\S]*?)\n\$\$)/gm, (match) =>

It looks for patterns like

> $$
something multiline
$$

I suggest to add also

text.replace(/(^> \$\$\n([\s\S]*?)\n> \$\$)/gm, (match) =>

to catch

> $$
> 3 \times g \times h_{\text{per group}} \times s^2 + d^2
> $$

and then remove > symbols in caught text (assumes LaTeX do not use them, maybe safer to remove only at start of the lines).

/^>//g - regex. do not know how to use it in html.

Does it makes sense?

Added:

To reduce code, might be beneficial to replace previous regex with text.replace(/(^> \$\$\n([\s\S]*?)\n[>]{0,1}[ ]{0,1}\$\$)/gm, (match) => (optional > ) and remove > symbols from start of lines. I guess [>]{0,1}[ ]{0,1} can/need to be replaced with >? ? for html code to work. Another variant (> )? is cleaner if works.

Also seems your code in the commit catches LaTeX with leaning >, so you probably already remove > ?