Package Details: llama.cpp-sycl b10048-1

Git Clone URL: https://aur.archlinux.org/llama.cpp-sycl.git (read-only, click to copy)
Package Base: llama.cpp-sycl
Description: llama.cpp with Intel Arc GPU acceleration via SYCL/oneAPI. Please read the README on GitHub before use.
Upstream URL: https://github.com/cantosun99/llama.cpp-sycl
Keywords: arc intel llama llama.cpp oneapi openvino sycl vulkan
Licenses: MIT
Conflicts: intel-oneapi-basekit-2025, intel-oneapi-toolkit, llama.cpp, llama.cpp-bin, llama.cpp-bin-noavx, llama.cpp-clblas-git, llama.cpp-clblast, llama.cpp-cublas-git, llama.cpp-cuda, llama.cpp-cuda-essentials-only, llama.cpp-cuda-git, llama.cpp-gfx1151, llama.cpp-git, llama.cpp-hip, llama.cpp-hip-gfx1151, llama.cpp-hipblas-git, llama.cpp-opencl, llama.cpp-openvino, llama.cpp-sycl-f16, llama.cpp-sycl-f16-git, llama.cpp-sycl-f32, llama.cpp-sycl-f32-git, llama.cpp-vulkan, llama.cpp-vulkan-bin, llama.cpp-vulkan-gemma4, llama.cpp-vulkan-git
Provides: llama.cpp
Submitter: cantosun99
Maintainer: cantosun99
Last Packager: cantosun99
Votes: 1
Popularity: 0.35
First Submitted: 2026-05-06 09:00 (UTC)
Last Updated: 2026-07-16 19:42 (UTC)

Required by (13)

Sources (1)

Latest Comments

1 2 Next › Last »

cantosun99 commented on 2026-07-11 07:01 (UTC)

If you encounter issues with the intel-deep-learning-essentials while installing this package, 99% of the time it's one of the following options:

  • If you previously or currently have the intel-oneapi-toolkit, intel-oneapi-basekit-2025 or intel-oneapi-hpckit packages installed, please uninstall them and verify that everything in /opt/intel is deleted. Most likely you don't even need everything that's included in the full Intel® oneAPI Toolkit for developers, that's why the Intel® Deep Learning Essentials for users exists. I suggest replacing the three packages with this one unless you are one of the few people that need the full kit.

  • If you are updating this package and it fails, you have to delete /home/yourusername/intel and /home/yourusername/.intel and retry the update. Please verify just in case, but there usually is nothing important in there anyway.

Please visit https://github.com/cantosun99/intel-deep-learning-essentials and https://github.com/cantosun99/llama.cpp-sycl to stay up to date.

cantosun99 commented on 2026-07-05 11:05 (UTC)

Important: SYCL now supports MTP and beats llama.cpp-vulkan in PP by ~150 tok/s while only being ~1.5 tok/s slower in TG.

Also SYCL has gotten a new release on the 1st of July with the Intel Intel® Deep Learning Essentials now being on version 2026.1.0 and the Intel® Deep Neural Network Library on 2026.0.1 so I expect the future llama.cpp releases to further optimize the speed of the SYCL backend.

Benchmarks ran with llama.cpp b9873 on a Sparkle Intel Arc Pro B70 passive using Unsloth's Qwen3.6 27B Q6_K on Arch Linux x86_64 with the Linux 7.1.2-3-cachyos kernel.

Since llama-bench doesn't support MTP, I used my recommended daily-usage flags, warmed the model up with "Say hi." and then benchmarked it for prompt processing with "Read the following text and then only reply with "ok" once you have done so." followed by the German Wikipedia entry for Black Sabbath converted to text. For token generation I also warmed it up with "Say hi." and benchmarked it with "Write a complete Snake game in a single Python file. Output only the code, no explanations, no commentary. Include: arrow key controls, collision detection, food spawning, score display, and a game over screen. Make it actually runnable." with thinking off.

Backend MTP PP TG

Vulkan off 317.32 tok/s 16.92 tok/s

Vulkan on 313.88 tok/s 33.40 tok/s

SYCL off 485.18 tok/s 15.75 tok/s

SYCL on 423.76 tok/s 31.64 tok/s

cantosun99 commented on 2026-05-18 12:55 (UTC)

Actually, I realized that the way the Intel installer works, the two packages will not be able to coexist together so I'll have to bundle the DNN into the DLE which used to be the way it was shipped anyway up until three weeks ago. Otherwise they just conflict with eachother and wouldn't even be removable by pacman anyway since they are both written into the same folder. So yeah, I made the two packages too soon, the seperate oneDNN has to go, my bad.

cantosun99 commented on 2026-05-18 11:40 (UTC)

So, I hope that this finally fixes everything you were criticising. Again, I'm just trying to do my best and learn from my mistakes.

Now I added the Intel DLE and DNN as two seperate packages and let the llama.cpp-sycl package use them as dependencies. The two packages have the 2026.0.0 as their version number and the llama.cpp-sycl now has the used llama.cpp release tag as the version number since it now doesn't always get the latest release (I misread what you meant when you criticised the package building against head, apologies for that), just like the other builds, llama.cpp-cuda or llama.cpp-vulkan do as well.

I know you will point out that a version of oneDNN already exists, but it pulls directly from the most current GitHub release and for this specific use-case it's absolutely critical, that both the DLE and DNN come directly from Intel in the same release (2026.0.0 at the time of writing) to retain compatibility. I think this is a valid reason to make another DNN package as what it actually provides, more specifically WHERE it gets it from, is vastly different and required for this to work.

I hope this finally fixes every complaint that you had, hcartiaux, and is a valid reason enough for these now three packages. Everything else like the intel-compute-runtime or the levelzero stuff isn't critical and I used the existing packages before, but these two HAVE to be these specific versions. And again, I do think that it's a valid goal to have the install be 10 GB less than the full kit, especially considering the costs of drives nowadays and having the specific llama.cpp version for Intel GPUs being up to date by someone that actually uses it daily.

hcartiaux commented on 2026-05-17 09:29 (UTC) (edited on 2026-05-17 09:30 (UTC) by hcartiaux)

Again, this is my first AUR package and I hope this now confirms with all the rules, if not, please let me know, I'm still learning and willing to improve!

My advice, start with something simpler, actually read the documentation on the wiki and don't use a LLM blindly, especially when answering another fellow human being.

Install the oneAPI toolchain temporarily during the build, then package it to /opt/intel/oneapi/

You try to save disk space using the wrong method. You should use makedepends, so that the build dependencies can be removed when running this package. You should not bundle anything related to the intel toolchain into this package, it's another thing, especially:

  • onednn is packaged in extra, no need to repackage it
  • Intel Deep Learning Essentials should be packaged separately, if it's not packaged already in extra or AUR. In example, there's already intel-oneapi-mkl-sycl in extra.

I would understand if you're willing to spend more time to split the required components included in intel-oneapi-basekit into smaller packages, but don't hide it in llama.cpp-sycl, it's simply not correct.

In the current form, you should probably create a container and put your installation commands in a Dockerfile instead of a PKGBUILD.

Clone llama.cpp from source (always builds against the latest HEAD)

You have ignored my remark here.

Honestly, I think that a maintained version of llama.cpp-sycl is nice to have in AUR, but I think this PKGBUILD is going in the wrong direction.

cantosun99 commented on 2026-05-17 08:09 (UTC)

Alright, I adressed hcartiaux's concerns and updated the package to conform with the AUR rules. Adressing each point of your concerns: 1. Instead of providing the extracted oneAPI install on my GitHub, the package now sources the offline-installer directly from Intel, extracts it and then builds from there. 2. Therefore you don't have to trust my tar files from my GitHub release, it how sources directly from Intel and even checks with the Intel-provided SHA384. 3. Now this package sources everything upsteam and changes nothing downstream I hope this makes the package confirm with the AUR rules and makes it as transparent as it can be. I still think saving close to 10GB of hard-drive space is a huge benefit over using the intel-oneapi-basekit and considering that kit is out of date, me using this package personally and thus keeping it up to date will also fix that problem. Again, this is my first AUR package and I hope this now confirms with all the rules, if not, please let me know, I'm still learning and willing to improve!

To sum up, copied from my README that I updated a minute ago: "Now the PKGBUILD will: 1. Download the Intel Deep Learning Essentials and oneDNN installers directly from Intel (pinned to a specific version, updated manually per release) 2. Install the oneAPI toolchain temporarily during the build, then package it to /opt/intel/oneapi/ 3. Clone llama.cpp from source (always builds against the latest HEAD) 4. Build it with SYCL enabled using Intel's icx/icpx compilers 5. Install shared libraries to /opt/llama.cpp-sycl/lib/ with RPATH baked in, keeping them out of the global /usr/lib namespace 6. Install binaries to /opt/llama.cpp-sycl/bin/ 7. Symlink all binaries into /usr/bin/ so they are accessible system-wide without any environment setup"

hcartiaux commented on 2026-05-16 14:23 (UTC)

Thanks for your contribution, I appreciate the efforts but I have huge concerns:

Arch Linux defines simplicity as without unnecessary additions or modifications. It ships software as released by the original developers—upstream—with minimal distribution-specific downstream changes.

You decrease the security of the package to save the cost of installing intel-oneapi-basekit. I really don't like it, I'm in favor of removing this package.

1 2 Next › Last »