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: 2.03
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 Next › Last »

siavoshkc commented on 2026-06-08 14:41 (UTC) (edited on 2026-06-08 14:42 (UTC) by siavoshkc)

The env was empty and it didn't work. But -DGGML_CCACHE=OFF in PKGBUILD fixed it. @Bink

Bink commented on 2026-06-08 13:59 (UTC) (edited on 2026-06-08 14:03 (UTC) by Bink)

Perhaps a misunderstanding @siavoshkc, but what I was giving you was an entire command.

$ CMAKE_CUDA_COMPILER_LAUNCHER= makepkg -sri
==> Making package: llama.cpp-cuda-git b9554.r2.19bba67c1f-1
==> Checking runtime dependencies...
==> Checking buildtime dependencies...
==> Retrieving sources...
  -> Cloning llama.cpp git repo...

...etc.

I'm not sure if it'll fix it, it's not an issue I've replicated, it's just something to try. I'm not using sccache.

The output of this might be helpful:

echo $CMAKE_CUDA_COMPILER_LAUNCHER

If it's not outputting sccache for you, then the proposed fix above is unlikely to help.

To try your way though, perhaps do this?

    local _cmake_options=(
    -G Ninja
    -B build
    -DCMAKE_CUDA_COMPILER_LAUNCHER=""
    -S "${_pkgname}"

siavoshkc commented on 2026-06-08 13:47 (UTC) (edited on 2026-06-08 13:49 (UTC) by siavoshkc)

@Bink I did this:

    local _cmake_options=(
    -G Ninja
    -B build
    -CMAKE_CUDA_COMPILER_LAUNCHER= makepkg -sri
    -S "${_pkgname}"

But got this:

  ==> Starting build()...
CMake Warning:
  Ignoring extra path from command line:

   "makepkg"


CMake Error: Unknown argument -sri
CMake Error: Run 'cmake --help' for all supported options.
==> ERROR: A failure occurred in build().

Besides, makepkg is what I run to build it like 'makepkg -Csi'. How telling cmake to use it as a compilation command can help?

Bink commented on 2026-06-08 12:43 (UTC) (edited on 2026-06-08 12:47 (UTC) by Bink)

@siavoshkc, if you're able to confirm if this is a viable workaround, I could potentially include CMAKE_CUDA_COMPILER_LAUNCHER in the PKGBUILD as a cmake variable, to override any local conflicting setting, for this build.

CMAKE_CUDA_COMPILER_LAUNCHER= makepkg -sri

siavoshkc commented on 2026-06-08 07:45 (UTC)

@Bink. sccache seems to be the issue. I renamed its executive and built manually from source. Don't know anything about how or where it is.

Bink commented on 2026-06-08 07:20 (UTC)

@siavoshkc, that looks like it might be an anomaly of your build environment. It's trying to use sccache, and failing. The PKGBUILD makes no requirement of sccache, but your environment has it defined for use.

Look at the output of:

echo $CMAKE_CUDA_COMPILER_LAUNCHER

Does that report sccache?

If so, try building with:

CMAKE_CUDA_COMPILER_LAUNCHER= makepkg -sri

(or a yay or paru command instead of makepkg -sri if that's being used)

Alternatively, you could build it in a clean chroot environment where sccache isn't used, but that may miss some of the GPU architecture specific optimisations.

siavoshkc commented on 2026-06-08 05:45 (UTC)

I see a bunch of these error: - FAILED: [code=254] ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o sccache /opt/cuda/bin/nvcc -forward-unknown-to-host-compiler -ccbin=/usr/bin/g++-15 -DGGML_BACKEND_BUILD -DGGML_BACKEND_SHARED -DGGML_CUDA_FA_ALL_QUANTS -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_USE_GRAPHS -DGGML_SCHED_MAX_COPIES=4 -DGGML_SHARED -D_GNU_SOURCE -D_XOPEN_SOURCE=600 -Dggml_cuda_EXPORTS -I/home/siavoshk/.cache/paru/clone/llama.cpp-cuda-git/src/llama.cpp/ggml/src/ggml-cuda/.. -I/home/siavoshk/.cache/paru/clone/llama.cpp-cuda-git/src/llama.cpp/ggml/src/../include -isystem /opt/cuda/targets/x86_64-linux/include -isystem /opt/cuda/targets/x86_64-linux/include/cccl -O3 -DNDEBUG -std=c++17 "--generate-code=arch=compute_86,code=[sm_86]" -Xcompiler=-fPIC -use_fast_math -extended-lambda -compress-mode=speed -Xcompiler "-Wmissing-declarations -Wmissing-noreturn -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-array-bounds -Wextra-semi -Wno-pedantic" -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o -MF ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o.d -x cu -c /home/siavoshk/.cache/paru/clone/llama.cpp-cuda-git/src/llama.cpp/ggml/src/ggml-cuda/argsort.cu -o ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o fatbinary fatal : Could not open input file 'argsort.cubin'

Bink commented on 2026-05-25 03:33 (UTC) (edited on 2026-05-25 03:34 (UTC) by Bink)

@jamesb6626, you're right, upstream did fix that in the very next commit after I'd pushed the update. I've commented it out for now and will likely remove the line altogether later.

Thanks again!

jamesb6626 commented on 2026-05-22 11:18 (UTC) (edited on 2026-05-22 11:19 (UTC) by jamesb6626)

@Bink, awesome, thanks. By the way, I think the llama.cpp commit bbce619 fixes the install for impl libraries to be automatic. I think this makes the added install line for helper libraries in package() redundant, though of course they should remain harmless probably.

Bink commented on 2026-05-22 02:15 (UTC)

@jamesb6626, thanks for bringing my attention to that. I've now implemented the change you suggested, and it's now correctly reporting the version in llama-server --version.

« First ‹ Previous 1 2 3 4 5 Next › Last »