[Bug]: Auto-scroll too sticky when LLM streams code

[ServeurpersoCom]

Browser

All

Description

When the LLM streams long code blocks, the chat view sticks so hard to the bottom that it's nearly impossible to scroll up and review the beginning of the conversation.

For normal text it feels fine ! but for code outputs it becomes frustrating: as soon as I try to scroll up, the view snaps back down.

Steps to Reproduce

Ask the LLM to spit out a big fat chunk of C++ code (no need for long lines ; it’s actually worse with short ones).
Try to scroll up, on PC with mousewhell, scrollbar or smartphone problem look the same.
Use fast >40 t/s generation on a small model or MoE to be sure

Expected Behavior

Auto-scroll should follow generation only if the user stays near the bottom.

If the user scrolls up past a threshold, auto-scroll should pause.

If the user scrolls back down to the bottom, auto-scroll can resume.

Version

No response

Additional Context

This is not trivial; I'm specifically looking for a good algorithm here. It's tricky to implement
The goal is to find a smooth balance between:
staying in sync with generated tokens,
and letting the user freely scroll without being pulled down.

[olegshulyakov]

Noticed it too, thinking on how to fix it.

[olegshulyakov]

@ServeurpersoCom Can you please check on latest version?

[ServeurpersoCom]

Git branch double-checked on main (it’s testing-branch5 on my side), browser caches synchronized. I tested 3 input methods on 2 devices:

Mouse Wheel on PC Browser ->The only one where I can kind of make it work, but my mouse has a free-spinning metal wheel so it goes flying ->it almost feels like cheating. Hard to say if there’s any real improvement.

Scrollbar on PC Browser ->Always get the ghost double-scroll effect. I’m on a 240 Hz monitor and the model is dense, spitting out ~50 tokens/second.

Mobile Touch (Samsung S25 Ultra) ->Also 120 Hz, but completely impossible.

It’s not trivial. Same behavior on regular work PCs I tested (60 Hz, same LLM).

NOTE: it really needs to be inside a code block plain text seems to go through!

Bug-Scroll-PC-AVC500kbps.mp4

Bug-Scroll-Mobile-AVC2Mbps.mp4

[ServeurpersoCom]

(Please remember to remove "invalid" from #178 ; Look at Claude, GPT-4, or anyone and send code to the LLM to compare with llama.ui Markdown must be displayed as plain text on the user side, and interpreted on the agent side )

[ServeurpersoCom]

On last master, now, I don't get the problem anymore, and it's fast !!!!!