ui: reduce per-token render cost when streaming

[zachwinter]

Overview

This pull request addresses rendering performance issues I identified in the llama-server UI. This builds on top of the existing work in #25225. I have an additional PR ready, pending this one.

All millisecond measurements are per streamed token.

e8a65bc adds a performance harness, used to ground the remaining work
6468958 CollapsibleContentBlock / CollapsibleTerminalBlock: 210.36ms -> 2.67ms
5346788 MarkdownContent + utils/code.ts: 11.58ms -> 0.62ms
da06dca LaTeX protection: 22.02ms -> 3.33ms
1023641 conversation store update-in-place: 3.07ms -> 1.36ms (@ 40 messages)

Additional information

This is a pure UI fix (no .cpp).

Requirements

• I have read and agree with the contributing guidelines: YES, and I'm still anxious I missed something (plz tell me).
• AI usage disclosure: I drove Opus for these fixes; I've been a UI Engineer by trade since 2009. Happy to remove the code comments if y'all would prefer it.

[ggml-gh-bot]

Hi @zachwinter, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• Large PR: Large changes require prior discussion (e.g. an issue or RFC) and maintainers may not be able to review this PR as-is. Consider splitting it into smaller, focused PRs.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[allozaur]

hey, thanks for this contribution! i had it planned to address this shortly, so that's a good timing :D lemme check it and get back to u

[ServeurpersoCom]

I'll take a look at that too

[allozaur]

@zachwinter i think the comments are good to stay for now, i think we will still re-visit some parts of how we render the UI so it's good for context.

Please merge my PR to your branch with few improvements — zachwinter/pull/1

[allozaur]

@zachwinter i think the comments are good to stay for now, i think we will still re-visit some parts of how we render the UI so it's good for context.

Please merge my PR to your branch with few improvements — zachwinter/pull/1

actually, closing this and will post a follow-up PR

[ServeurpersoCom]

LGTM to rework with Allozaur (follow-up)

[zachwinter]

@zachwinter i think the comments are good to stay for now, i think we will still re-visit some parts of how we render the UI so it's good for context.

Please merge my PR to your branch with few improvements — zachwinter/pull/1

actually, closing this and will post a follow-up PR

aye! apologies, I missed your last comments.

[ServeurpersoCom]

To be verified: the load with 256K tokens as well

[zachwinter]

@ServeurpersoCom the rendering performance should be independent from token count at this point – that was my north star. empiricism always wins, though!

[allozaur]

@zachwinter i guess let's proceed regardless, i will reference this PR in the follow-up one

[zachwinter]

@zachwinter i guess let's proceed regardless, i will reference this PR in the follow-up one

hard to understate my excitement. hashtag first merge!

I will tee up my second round shortly!

[allozaur]

@zachwinter i guess let's proceed regardless, i will reference this PR in the follow-up one

hard to understate my excitement. hashtag first merge!

I will tee up my second round shortly!

no worries i'm pushing from my branch and will post a PR in a moment

[zachwinter]

@ServeurpersoCom the rendering performance should be independent from token count at this point – that was my north star. empiricism always wins, though!

@ServeurpersoCom sorry, this comment is true, given this: #26098

[ServeurpersoCom]

@ServeurpersoCom the rendering performance should be independent from token count at this point – that was my north star. empiricism always wins, though!

@ServeurpersoCom sorry, this comment is true, given this: #26098

cc @allozaur

I’ll do a full review at the end: I approved it without benchmark (only a non regression quick test with a prompt like "Create a long, feature-rich Markdown demo to stress-test my streaming Markdown rendering engine. You never stop!!! Seriously, never! Throw in some KaTeX, too like a crazy math teacher!!!!") to let allozaur move forward, but I’ll conduct a comprehensive analysis at the end regardless.