Package Details: llama.cpp-cuda-git b9893.r0.6f8895feec-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: llama.cpp
Provides: llama.cpp
Submitter: Bink
Maintainer: Bink
Last Packager: Bink
Votes: 6
Popularity: 2.15
First Submitted: 2026-01-08 09:17 (UTC)
Last Updated: 2026-07-07 05:16 (UTC)

Dependencies (17)

Required by (13)

Sources (3)

Latest Comments

« First ‹ Previous 1 2 3 4 5

tixwho commented on 2026-04-23 03:30 (UTC)

Hi, I have a single Nvidia GPU on the device. I can build the package without issue until version b8892.r0.0d0764dfd2-1, but now the package failed to build for llama-server. The error log is as follows:

[546/546] Linking CXX executable bin/llama-server
FAILED: [code=1] bin/llama-server 
: && /usr/bin/c++ -march=x86-64 -mtune=generic -O2 -pipe -fno-plt -fexceptions         -Wp,-D_FORTIFY_SOURCE=3 -Wformat -Werror=format-security         -fstack-clash-protection -fcf-protection         -fno-omit-frame-pointer -mno-omit-leaf-frame-pointer -Wp,-D_GLIBCXX_ASSERTIONS -g -ffile-prefix-map={user}/.cache/paru/clone/llama.cpp-cuda-git/src=/usr/src/debug/llama.cpp-cuda-git -flto=auto -O3 -DNDEBUG -Wl,-O1 -Wl,--sort-common -Wl,--as-needed -Wl,-z,relro -Wl,-z,now          -Wl,-z,pack-relative-relocs -flto=auto -Wl,--dependency-file=tools/server/CMakeFiles/llama-server.dir/link.d tools/server/CMakeFiles/llama-server.dir/server.cpp.o tools/server/CMakeFiles/llama-server.dir/server-http.cpp.o tools/server/CMakeFiles/llama-server.dir/server-models.cpp.o -o bin/llama-server  -Wl,-rpath,{user}/.cache/paru/clone/llama.cpp-cuda-git/src/build/bin:  tools/server/libserver-context.a  bin/libllama-common.so.0.0.0  vendor/cpp-httplib/libcpp-httplib.a  common/libllama-common-base.a  bin/libmtmd.so.0.0.0  bin/libllama.so.0.0.0  bin/libggml.so.0.10.0  bin/libggml-cpu.so.0.10.0  bin/libggml-cuda.so.0.10.0  /opt/cuda/targets/x86_64-linux/lib/stubs/libcuda.so  bin/libggml-rpc.so.0.10.0  bin/libggml-base.so.0.10.0  /usr/lib/libssl.so  /usr/lib/libcrypto.so && :
/usr/bin/ld: final link failed: bad value
collect2: error: ld returned 1 exit status
ninja: build stopped: subcommand failed.
==> ERROR: A failure occurred in build().
    Aborting...
error: failed to build 'llama.cpp-cuda-git-b8851.r0.e365e658f0-1': 
error: packages failed to build: llama.cpp-cuda-git-b8851.r0.e365e658f0-1

Installing the two optional dependencies nccl and ccache didn't fix the issue. I wonder if it could be due to some recent build flag changes?

TeddyHuang-00 commented on 2026-04-22 19:55 (UTC) (edited on 2026-04-22 19:55 (UTC) by TeddyHuang-00)

@dareme7 I am also using fish shell. The only two things cuda.sh does is 1. Add /opt/cuda/bin to path, and 2. set NVCC_CCBIN to /usr/bin/g++.

You can bypass the script by setting NVCC_CCBIN to the same /usr/bin/g++ (PKGBUILD will be happy with it set), and of course add /opt/cuda/bin to your path using fish_add_path to make sure nvcc is available.

I don't know if it is the best thing to do to put them into your config, but should definitely get the issues resolved for now.

dareme7 commented on 2026-04-21 20:52 (UTC)

This now crashes by saying append_path command in /etc/profile.d/cuda.sh is not found.

By googling I found a forum thread saying PKGBUILD should not be sourcing like this.

For reference I'm using the fish shell, which might be relevant here.

Bink commented on 2026-04-20 01:25 (UTC) (edited on 2026-05-22 02:21 (UTC) by Bink)

The package now leverages ninja to ensure parallel builds regardless of makepkg.conf settings.

Builds native for GGML and CUDA by default, for improved optimisation. Use the aur_llamacpp_build_universal variable to produce a build for all CPU/CUDA-architecture variants, if needed. For example, to target universal architecture:

aur_llamacpp_build_universal=true makepkg -sri

If you have multiple Nvidia GPU's, be sure to install the optional dependency nccl, for multi-GPU parallelism.

To improve re-build times, install the optional dependency ccache.

Bink commented on 2026-02-23 22:29 (UTC) (edited on 2026-02-23 22:32 (UTC) by Bink)

Hi @zeroxoneafour. Thanks for the feedback! Suggestions are always welcome!

The -j$(nproc) flag should be added to your /etc/makepkg.conf file, and you'll then enjoy the benefit of that optimisation for all applicable compile tasks.

See here: https://wiki.archlinux.org/title/Makepkg#Parallel_compilation

The Release build is already specified in the PKGBUILD:

-DCMAKE_BUILD_TYPE=Release

zeroxoneafour commented on 2026-02-23 05:44 (UTC)

Should add --config Release -j $(nproc) to the cmake build to speed up build times.

Bink commented on 2026-01-30 01:16 (UTC)

Thanks for spotting and sharing that @ryant0000. I've now removed that flag.

ryant0000 commented on 2026-01-24 17:42 (UTC)

CMake Warning at CMakeLists.txt:121 (message): LLAMA_CURL option is deprecated and will be ignored

This can/should probably be removed.

« First ‹ Previous 1 2 3 4 5