Package Details: llama.cpp-cuda-git b11433.r0.50569eb87d-1

Git Clone URL: https://aur.archlinux.org/llama.cpp-cuda-git.git (read-only, click to copy)
Package Base: llama.cpp-cuda-git
Description: Port of Facebook's LLaMA model in C/C++ (with NVIDIA CUDA optimizations)
Upstream URL: https://github.com/ggml-org/llama.cpp
Licenses: MIT
Conflicts: ggml, ggml-cuda-git, libggml, llama.cpp
Provides: ggml, ggml-cuda-git, libggml, libggml-base.so, libggml-cpu.so, libggml-cuda.so, libggml-rpc.so, libggml.so, libllama-batched-bench-impl.so, libllama-bench-impl.so, libllama-cli-impl.so, libllama-common.so, libllama-completion-impl.so, libllama-fit-params-impl.so, libllama-perplexity-impl.so, libllama-quantize-impl.so, libllama-server-impl.so, libllama.so, libmtmd.so, llama.cpp
Replaces: ggml-cuda-git
Submitter: Bink
Maintainer: Bink
Last Packager: Bink
Votes: 7
Popularity: 1.27
First Submitted: 2026-01-08 09:17 (UTC)
Last Updated: 2026-10-05 23:38 (UTC)

Dependencies (17)

Required by (16)

Sources (3)

Pinned Comments

Bink commented on 2026-09-26 12:25 (UTC)

This is a VCS (-git) package. If you want to build and install from the latest commit, simply re-build the package, pulling in the latest commit. Don't flag this package out-of-date if it's otherwise building fine. If build is failing, please let me know.

VCS packages are not considered out of date when the pkgver changes; do not flag them as the maintainer will merely unflag the package and ignore you. AUR maintainers should not commit mere pkgver bumps.

Arch Wiki - AUR - Flagging packages out-of-date

Latest Comments

1 2 3 4 5 6 Next › Last »

Bink commented on 2026-09-26 12:25 (UTC)

This is a VCS (-git) package. If you want to build and install from the latest commit, simply re-build the package, pulling in the latest commit. Don't flag this package out-of-date if it's otherwise building fine. If build is failing, please let me know.

VCS packages are not considered out of date when the pkgver changes; do not flag them as the maintainer will merely unflag the package and ignore you. AUR maintainers should not commit mere pkgver bumps.

Arch Wiki - AUR - Flagging packages out-of-date

Bink commented on 2026-09-14 22:54 (UTC)

Thanks @ddd! I've tested, confirmed and pushed the update ;)

ddd commented on 2026-09-14 14:04 (UTC)

CUDA now supports GCC 16, gcc15 can be dropped from makedepends

siavoshkc commented on 2026-09-11 16:30 (UTC) (edited on 2026-09-11 16:31 (UTC) by siavoshkc)

When I install the package I have to do sudo mv /usr/bin/sccache /usr/bin/sccache.disabled otherwise I get 

[30/523] Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.o
FAILED: [code=254]
...
fatbinary fatal   : Could not open input file 'argsort.cubin'

Bink commented on 2026-09-07 04:45 (UTC) (edited on 2026-09-07 04:47 (UTC) by Bink)

The dependency on the separate ggml-cuda-git package has been removed, as whisper-cpp-cuda-git now leverages llama.cpp functionality, and can therefore depend on this package instead.

This naturally resolves the doubling-up issue where it was pulling the same source twice, fairly noted by some.

void09 commented on 2026-08-06 18:54 (UTC)

@Bink actually my connection is fast, 2.5gbit ~300MB/s. pulling this package from git took more than 5 minutes, close to 10. I noticed it is faster on my other 1g isp, so a peering issue might also be involved, but the package is still slow to get. Thanks for looking into it, I will see if there is a way to make it faster without breaking the precious standards.

Bink commented on 2026-08-05 05:28 (UTC) (edited on 2026-08-05 05:29 (UTC) by Bink)

@void09, I can understand your need, particularly on a slower connection to try and eliminate the doubling up of that downloaded source. What you shared however, would step on the toes of a number Arch packaging standards.

What I'd like to suggest instead, is setting the SRCDEST variable within your /etc/makepkg.conf. This sets a common place for downloaded sources, that both packages would then share.

Something like this in /etc/makepkg.conf (you can probably just un-comment the existing line):

SRCDEST=/home/sources

void09 commented on 2026-08-04 18:22 (UTC)

I noticed this package takes quite a long time to pull from git, on 2g connection. I am not good with arch packages, so I asked an ai agent to investigate. It seems that you are pulling the same repo twice, once for the ggml and once for the llamcpp, and they do not get cached, since they are different packages. https://paste.rs/2QxjM https://paste.rs/8Cosc suggested these changes, maybe you can look it.

siavoshkc commented on 2026-07-09 19:29 (UTC)

ggml-cuda-git-1:b9940.r9.049326a000-1 and llama.cpp-cuda-git-b9862.r0.5a460dea9f-1 are in conflict

Bink commented on 2026-07-08 01:55 (UTC) (edited on 2026-07-08 01:59 (UTC) by Bink)

Thanks for bringing that issue to my attention @rubin55.

Just a few hours ago, a new commit to llama.cpp bumped the llama.cpp and ggml codebases out of alignment, causing that issue.

That could happen again in the future too, so I've pushed an update to the ggml-cuda-git dependency so it's now pulling and building GGML from upstream llama.cpp/ggml instead, which should ensure it remains aligned with this package, so long as both this package, and ggml-cuda-git are re-built together.

1 2 3 4 5 6 Next › Last »