[Submission] llama-cpp-bin

aminya · 2026-10-13 10:38:47

Hi, I have created the llama-cpp-bin AUR packages for all variants, I'd like to submit it. Since the AUR registration is currently closed, I’m submitting this through the forum.
https://github.com/aminya/llama-cpp-bin.git

| Variant   | Stable package            | Prerelease package                   |
| --------- | ------------------------- | ------------------------------------ |
| CPU       | `llama-cpp-bin`           | `llama-cpp-prerelease-bin`           |
| Vulkan    | `llama-cpp-vulkan-bin`    | `llama-cpp-vulkan-prerelease-bin`    |
| CUDA 12   | `llama-cpp-cuda12-bin`    | `llama-cpp-cuda12-prerelease-bin`    |
| CUDA 13   | `llama-cpp-cuda13-bin`    | `llama-cpp-cuda13-prerelease-bin`    |
| ROCm      | `llama-cpp-rocm-bin`      | `llama-cpp-rocm-prerelease-bin`      |
| OpenVINO  | `llama-cpp-openvino-bin`  | `llama-cpp-openvino-prerelease-bin`  |
| SYCL FP16 | `llama-cpp-sycl-fp16-bin` | `llama-cpp-sycl-fp16-prerelease-bin` |
| SYCL FP32 | `llama-cpp-sycl-fp32-bin` | `llama-cpp-sycl-fp32-prerelease-bin` |

Last edited by aminya (2026-10-13 10:40:48)

cryptearth · 2026-10-13 11:47:14

why?
there's https://archlinux.org/packages/extra/x86_64/llama-cpp/
and https://wiki.archlinux.org/title/Llama.cpp explains the backends https://archlinux.org/packages/?sort=&q … =&flagged=
your proposal violates, in fact, "no duplicates"

Last edited by cryptearth (2026-10-13 11:47:44)

Lone_Wolf · 2026-10-13 20:00:17

Moderator Note

Moving to AUR Issues, Discussion & PKGBUILD Requests

aminya · 2026-10-14 07:50:30

cryptearth wrote:

why?
there's https://archlinux.org/packages/extra/x86_64/llama-cpp/
and https://wiki.archlinux.org/title/Llama.cpp explains the backends https://archlinux.org/packages/?sort=&q … =&flagged=
your proposal violates, in fact, "no duplicates"

The method you linked doesn't allow simultaneously installing different backends, or switching between them, and also, it does not provide pre-releases.

Lone_Wolf · 2026-10-14 08:43:05

The repo ggml-* backends don't list conflicts in their PKGBUILDs so should be installable together.
Have you tried if it is possible ?

The llama-cpp-prerelease-bin would probably be allowed in the aur, though it needs some changes .

example

conflicts=(
  'llama.cpp'
  'llama.cpp-bin'
  'llama.cpp-cuda'
  'llama.cpp-cuda-bin'
  'llama.cpp-cuda13-bin'
  'llama.cpp-hip'
  'llama.cpp-sycl-f32'
  'llama.cpp-vulkan'
  'llama.cpp-vulkan-bin'
  'llama-cpp'
  'llama-cpp-bin'
)

Don't try to conflict with every possible variant in existence, just conflict with the primary package (typically a sourcebuild from repos).
If there are conflicts that can't be detected/prevented by pacman the PKGBUILDs need to deal with them (preferably with an explanation)

cryptearth · 2026-10-14 18:48:44

aminya wrote:

The method you linked doesn't allow simultaneously installing different backends, or switching between them

you clearly have not tested that yourself

pacman -Ss ggml
extra/ggml 0.26.0-1 [installed]
    Tensor library for machine learning
extra/ggml-blas 0.26.0-1 [installed]
    Tensor library for machine learning
extra/ggml-cuda 0.26.0-1 [installed]
    Tensor library for machine learning
extra/ggml-hip 0.26.0-1 [installed]
    Tensor library for machine learning
extra/ggml-openvino 0.26.0-1 [installed]
    Tensor library for machine learning
extra/ggml-sycl 0.26.0-1
    Tensor library for machine learning
extra/ggml-vulkan 0.26.0-1 [installed]
    Tensor library for machine learning

the only hard conflict is between base ggml and ggml-sycl

, and also, it does not provide pre-releases.

which is important exactly WHY? for someone using arch and if they, for some reason, need the current git master one can expect any more or less competent arch user to get this done on thier own - at least i don'T see any need for it - and if you really need someone else to do THAT for you then you probably shouldn't use either arch or and pre-/-git releases anyway but only whats in your distros official repo
sorry - but failed AGAIN buddy
if you want to try to push something you should at least put some own effort into it - to me this has quite some "done BY a-not-i" written all over it
still voting for "violated rule: avoid duplicates"

// edit

github profile wrote:

Lead engineer and founder building AI

as if i knew it before reading it

jeez - please let the ai bubble finally pop to spare us from such nonesense

// edit 2

530 repos and "30.000+ contributions" from a guy that looks like Max Verstappen in his mid-20s ... yea ... forget about it - at least 99,9% AI slop for sure

Last edited by cryptearth (2026-10-14 19:00:35)

seth · 2026-10-14 20:06:30

Most of them are forks and he registered 2015 and gets active 2019 - well before the LLM noise.
I suspect pre-releases might also be interesting if you operate in that field and there're many subrelease packages, as long as they're unique that's not disqualifying.

Max Verstappen in his mid-20ies looked like the average 15yr old - except in much better shape

Arch Linux

#1 2026-10-13 10:38:47

[Submission] llama-cpp-bin

#2 2026-10-13 11:47:14

Re: [Submission] llama-cpp-bin

#3 2026-10-13 20:00:17

Re: [Submission] llama-cpp-bin

#4 2026-10-14 07:50:30

Re: [Submission] llama-cpp-bin

#5 2026-10-14 08:43:05

Re: [Submission] llama-cpp-bin

#6 2026-10-14 18:48:44

Re: [Submission] llama-cpp-bin

#7 2026-10-14 20:06:30

Re: [Submission] llama-cpp-bin

Board footer