Arch Linux User Repository

Alternate versions

llama.cpp
llama.cpp-vulkan
llama.cpp-sycl-fp16
llama.cpp-sycl-fp32
llama.cpp-cuda
llama.cpp-cuda-f16
llama.cpp-hip

I am able to build the package with manual modifications. Steps posted below:

1. Fetch resources and extract the files: makepkg -o
2. Manually build the package:

cd src
source /opt/intel/oneapi/setvars.sh
local _cmake_options=(
    -B build
    -S llama.cpp
    #-DCMAKE_BUILD_TYPE=None
    -DCMAKE_INSTALL_PREFIX='/usr'
    -DGGML_ALL_WARNINGS=OFF
    -DGGML_ALL_WARNINGS_3RD_PARTY=OFF
    #-DBUILD_SHARED_LIBS=OFF
    #-DGGML_STATIC=ON
    #-DGGML_LTO=ON
    -DGGML_RPC=ON
    -DLLAMA_CURL=ON
    -DGGML_BLAS=ON
    -DCMAKE_C_COMPILER=icx
    -DCMAKE_CXX_COMPILER=icpx
    -DGGML_SYCL=ON
    -DGGML_SYCL_F16=ON # Comment this out for building F32 version
    -Wno-dev
)
cmake "${_cmake_options[@]}"
cmake --build build --config Release -j -v

1. Patch the PKGBUILD, as there are no lib*.a in a non-static build.

diff --git a/PKGBUILD b/PKGBUILD
index 7999ec8..51cc1bc 100644
--- a/PKGBUILD
+++ b/PKGBUILD
@@ -73,7 +73,7 @@ build() {
 package() {
   DESTDIR="${pkgdir}" cmake --install build
   rm "${pkgdir}/usr/include/"ggml*
-  rm "${pkgdir}/usr/lib/"lib*.a
+  #rm "${pkgdir}/usr/lib/"lib*.a

   install -Dm644 "${_pkgname}/LICENSE" "${pkgdir}/usr/share/licenses/${pkgname}/LICENSE"

1. Package the build and install: makepkg -Ri

Notes:

1. BUILD_SHARED_LIBS and GGML_STATIC were commented out because of the inference issue. The built package will have libggml.so shared libraries which may have conflicts with other packages.
2. Switching GGML LTO off will fix build error libggml-sycl.a: error adding symbols: archive has no index; run ranlib to add one. I think it is some limitations of icx/icpx.
3. I've noticed some differences between build/compile_commands.json in manual build and pacman build, with the same build flags. Probably related to the SYCL_EXTERNAL issue.

--- /tmp/pacman.json    2025-03-16 10:41:15.426354158 +0800
+++ /tmp/manual.json    2025-03-16 10:41:06.154835418 +0800
@@ -1,31 +1,47 @@
 {
   "directory": "/home/heyrict/.cache/paru/clone/llama.cpp-sycl-f32/src/build/ggml/src",
   "command": "/opt/intel/oneapi/compiler/2025.0/bin/icx
  -DGGML_BUILD
  -DGGML_SCHED_MAX_COPIES=4
  -DGGML_SHARED
  -D_GNU_SOURCE
  -D_XOPEN_SOURCE=600
  -Dggml_base_EXPORTS
  -I/home/heyrict/.cache/paru/clone/llama.cpp-sycl-f32/src/llama.cpp/ggml/src/.
  -I/home/heyrict/.cache/paru/clone/llama.cpp-sycl-f32/src/llama.cpp/ggml/src/../include
+ -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
+ -g
+ -ffile-prefix-map=/home/heyrict/.cache/paru/clone/llama.cpp-sycl-f32/src=/usr/src/debug/llama.cpp-sycl-f32
+ -flto=auto
  -O3
  -DNDEBUG
  -std=gnu11
  -fPIC
  -Wshadow
  -Wstrict-prototypes
  -Wpointer-arith
  -Wmissing-prototypes
  -Werror=implicit-int
  -Werror=implicit-function-declaration
  -Wall
  -Wextra
  -Wpedantic
  -Wcast-qual
  -Wno-unused-function
  -o CMakeFiles/ggml-base.dir/ggml.c.o
  -c /home/heyrict/.cache/paru/clone/llama.cpp-sycl-f32/src/llama.cpp/ggml/src/ggml.c",
   "file": "/home/heyrict/.cache/paru/clone/llama.cpp-sycl-f32/src/llama.cpp/ggml/src/ggml.c",
   "output": "ggml/src/CMakeFiles/ggml-base.dir/ggml.c.o"
 },

Still broken: Current build log at https://abs-cd.oscloud.info/cd_manager/llama.cpp-sycl-f16

You get past the SYCL_EXTERNAL issue by providing a makepkg.conf without most of the -fsomething C(XX)FLAGS, but run into architecture problems afterwards.

intel-compute-runtime should be an optdepend of this and the -f32 pkg. It's needed for OneAPI GPU access.

@greyltc I could neither build it (with intel-oneapi-basekit-2025) in a clean buildenv, with a custom makepkg.conf that removes all the C(XX)FLAGS nor manually. I got past the undefined function without SYCL_EXTERNAL with the latter 2, but the got a shitton of errors related to typedefs where my first guess is that the compiler should have used stdlib headers under /opt/intel/oneapi. Currently trying to build in docker.

Edit: Docker builds & runs, but I didn't have any luck building it on Arch. I also tried changing the BLAS libary to OpenBLAS, setting -march=haswell to maybe get those weird intrinsincs errors away, but to no help.

I'm not sure the SYCL_EXTERNAL issue is upstream's problem. I think it might be caused by the environment variables that makepkg sets when it builds packages. Still trying to figure it out...

Thanks @grelytc! Could you post an upstream issue about SYCL_EXTERNAL?

@txtsd, you can depend on https://aur.archlinux.org/packages/intel-oneapi-basekit-2025 here if you want the latest intel-oneapi-basekit. That solves the "'syclcompat/math.hpp' file not found" issue, but even with that solved the package still won't build.

While this package cannot be build one may use Intel's ipex-llm project, that compatible with llama.cpp and works very fast with Archlinux intel-oneapi-basekit 2024.1.

Here is official guide: https://github.com/intel-analytics/ipex-llm/blob/main/docs/mddocs/Quickstart/llama_cpp_quickstart.md

I was told they wont update it until after the Python 3.13 migration.

I have updated it locally and am now getting

/usr/lib64/gcc/x86_64-pc-linux-gnu/14.2.1/../../../../include/c++/14.2.1/array:217:2: error: SYCL kernel cannot call an undefined function without SYCL_EXTERNAL attribute

It's the intel-oneapi-basekit package. It's been outdated for a long time now. I'll send a patch and see if they'll accept it.