Misc. bug: Web UI is gone (b9680)

[maikelthedev]

Name and Version

version: 9680 (d5376cf)
built with GNU 14.2.0 for Linux x86_64

Operating systems

Linux (NixOS 26.05)

Which llama.cpp modules do you know to be affected?

llama-server

Command line

podman run --rm --device nvidia.com/gpu=all \
    -p 8090:8090 \
    -v /home/maikel/models:/home/maikel/models:ro \
    -v /home/maikel/code/nixconf/llms/presets/llama-router-presets.ini:/home/maikel/code/nixconf/llms/presets/llama-router-presets.ini:ro \
    ghcr.io/ggml-org/llama.cpp:server-cuda13-b9680 \
    --host 0.0.0.0 --port 8090 --models-preset /home/maikel/code/nixconf/llms/presets/llama-router-presets.ini --models-max 1

Problem description & steps to reproduce

Web UI doesn't work at all. Any address, any server, they are all gone. Just JSON "file not found"

First Bad Commit

EDIT: I tried now and can confirm, it's b9680, the official doocker image server-cuda13-b9664 works flawlesly, whereas server-cuda13-b9680 has the problem above. Exact same command just swapping versions.

Relevant log output

n/a

[taronaeo]

Any error message from the container?

[ngxson]

seems like there was a problem downloading assets from HF: https://github.com/ggml-org/llama.cpp/actions/runs/27674244014/job/81857248406

not sure if we should exit with a status code error in this case, instead of failing silently - your thoughts? @CISC @aldehir @allozaur

[aldehir]

@ngxson This is how it first worked, not on purpose, but it failed hard and resulted in an influx of issues created.

My opinion is to use a multi-stage build and build the UI directly instead of pulling from HF. The HF downloading is IMO mostly for end user convenience, and doing it in CI is unstable because it relies on the ui-publish workflow to run first.

[vbooka1]

This seems to be a resurrection of the recent bug #23156 why the UI components were moved to HF in the first place?
Edit: sorry, I did not read the comment #24726 (comment)

[allozaur]

not sure if we should exit with a status code error in this case, instead of failing silently - your thoughts? @CISC @aldehir @allozaur

definitely shouldn't fail silently, exit with error status is the way to go. @aldehir can you please take over this one?

[segmond]

Yup, was working for me then started failing. I frankly don't want to run npm on this box. Too many supply chain attacks if you have been following trends lately.