llama.cpp-sycl

llama.cpp with Intel GPU acceleration, packaged for Arch Linux.

Intel makes great GPUs. They also make it nearly impossible to use them on anything that isn't Ubuntu. The oneAPI toolchain comes with apt, zypper, and yum/dnf support, nothing for Arch. The AUR has the full intel-oneapi-toolkit package which is gigabytes of stuff you'll never use to actually run llama.cpp. If you want SYCL acceleration on Arch, you're basically on your own.

This package fixes that. It builds llama.cpp on your machine with full SYCL support at close to 10 GB less than the full kit.

---

Current stats

SYCL got a lot of love recently and currently gets peaks of 37 tok/s TG and 900 tok/s PP with Qwen3.8 27B which beats llama.cpp-vulkan by a long shot.

With my current recommendations, the averages are roughly 27 TG while the model is thinking, 33 TG actually outputting/writing, 650-700 PP, tested with a Sparkle Intel Arc Pro B70 at 275W, using the Linux 7.1.8-1-cachyos kernel.

Current news

(August 20th) Unsloth released new Unsloth Dynamic v3.0 quants for Qwen3.8-27B which currently are my recommendation. If your gguf is older than August 19th, you should re-download it. I have updated the recommendations with two examples.

Example 1: UD-Q6_K_XL, the highest quant you can run on a B70 for maximum quality with a small but usally sufficient context of 100000 tokens. Highest quality but still enough context to handle the xhigh reasoning for short tasks that require accuracy and attention to detail. This should be your "daily-driver" for 90% of tasks.

Example 2: UD-Q4_K_XL, the highest quant you can run on a B70 that still allows you to use the full 262144 token context. Slightly reduced quality but a huge amount of context, ideal for tasks that don't require as much accuracy or intensive reasoning and instead work with a ton of data.

(August 14th, midnight) I finally got done with a lot of testing and I have to say, Qwen3.8 27B is genuinely insane. Not one failed tool call, insanely time-consuming reasoning and very token-hungry, but in the end it's worth it because it perfectly one-shot most of the tasks. Updated my recommendations, off to bed now, have fun trying it out!

(August 11th) The AUR is finally back again!

---

What this does

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
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

---

Install

Via AUR helper

yay -S llama.cpp-sycl
# or
paru -S llama.cpp-sycl

Manual

git clone https://github.com/cantosun99/llama.cpp-sycl.git
cd llama.cpp-sycl
makepkg -si

The build takes a while depending on your CPU. This is normal. On my 270K it takes about two to three minutes and is about 6.0 GB installed in total.

For further information, you can visit the llama.cpp documentation of the SYCL backend https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/SYCL.md or the article by Intel https://www.intel.com/content/www/us/en/developer/articles/technical/run-llms-on-gpus-using-llama-cpp.html

---

First-time setup

After installation, verify that everything is working before running your first model. Versions may differ.

1. Use bash

The oneAPI setvars.sh script requires bash. If your default shell is fish or zsh, switch to bash:

bash

2. Load the oneAPI environment

source /opt/intel/oneapi/setvars.sh

Expected output:

:: initializing oneAPI environment ...
   bash: BASH_VERSION = 5.3.15(1)-release
   args: Using "$@" for setvars.sh arguments: 
:: ccl -- latest
:: compiler -- latest
:: debugger -- latest
:: dev-utilities -- latest
:: dnnl -- latest
:: dpl -- latest
:: mkl -- latest
:: mpi -- latest
:: pti -- latest
:: tbb -- latest
:: tcm -- latest
:: umf -- latest
:: oneAPI environment initialized ::

3. Verify the compiler

icpx --version

Expected output:

Intel(R) oneAPI DPC++/C++ Compiler 2026.1.0 (2026.1.0.20260617)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /opt/intel/oneapi/compiler/2026.1/bin/compiler
Configuration file: /opt/intel/oneapi/compiler/2026.1/bin/compiler/../icpx.cfg

4. Verify your GPU is detected by SYCL

sycl-ls

You should see at least one level_zero:gpu entry for your Intel GPU:

[level_zero:gpu][level_zero:0] Intel(R) oneAPI Unified Runtime over Level-Zero V2, Intel(R) Arc(TM) Pro B70 Graphics 20.2.0 [1.15.38646]
[level_zero:gpu][level_zero:1] Intel(R) oneAPI Unified Runtime over Level-Zero V2, Intel(R) Graphics 12.70.4 [1.15.38646]
[opencl:cpu][opencl:0] Intel(R) OpenCL, Intel(R) Core(TM) Ultra 7 270K Plus OpenCL 3.0 (Build 0) [2026.21.6.0.17_160000]
[opencl:gpu][opencl:1] rusticl, Mesa Intel(R) Graphics (BMG G31) OpenCL 3.0  [26.1.4-arch3.1]
[opencl:gpu][opencl:2] rusticl, Mesa Intel(R) Graphics (ARL) OpenCL 3.0  [26.1.4-arch3.1]
[opencl:gpu][opencl:3] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) Pro B70 Graphics OpenCL 3.0 NEO  [26.22.38646]
[opencl:gpu][opencl:4] Intel(R) OpenCL Graphics, Intel(R) Graphics OpenCL 3.0 NEO  [26.22.38646]

5. Verify llama.cpp sees the GPU

/opt/llama.cpp-sycl/bin/llama-cli --list-devices

You should see your GPU listed as a SYCL device:

Available devices:
  SYCL0: Intel(R) Arc(TM) Pro B70 Graphics (32656 MiB, 31665 MiB free)
  SYCL1: Intel(R) Graphics (29115 MiB, 11642 MiB free)

If all steps above produce output similar to the examples, you're ready to go.

---

Recommended usage

Every time you want to run llama.cpp, you need to load the oneAPI environment first, not just copy-paste your llama-server command as you would with other builds.

In my opinion Qwen3.8-27B is the best model you can currently run and it's not even remotely close. Gemma 4 and Qwen3.6 are outdated, Muse Glimmer has the dumbest relase date I've heard in my life, you can't just launch an inferior model to Qwen3.6 in the same week that Qwen3.8 releases. There might be use-cases for Nemotron 3.5 Lightning, if you depend on fast and accurate tool calls for example.

Unsloth's guide how to run Qwen3.8-27B Unslot's Hugging Face Repo for Qwen3.8-27B

Example 1: UD-Q6_K_XL, the highest quant you can run on a B70 for maximum quality with a small but usally sufficient context of 100000 tokens. Highest quality but still enough context to handle the xhigh reasoning for short tasks that require accuracy and attention to detail. This should be your "daily-driver" for 90% of tasks.

source /opt/intel/oneapi/setvars.sh
/opt/llama.cpp-sycl/bin/llama-server \
  --model /path/to/your/model/Qwen3.8-27B-UD-Q6_K_XL.gguf \
  --device SYCL0 \
  --n-gpu-layers 999 \
  --load-mode none \
  --flash-attn on \
  --jinja \
  --reasoning-preserve \
  --ctx-size 100000 \
  --cache-type-k q8_0 \
  --cache-type-v q8_0 \
  --temp 1.0 \
  --top-p 0.95 \
  --top-k 20 \
  --min-p 0.00 \
  --presence-penalty 0.0 \
  --repeat-penalty 1.0 \
  --spec-type draft-mtp \
  --spec-draft-n-max 2 \
  --port 9931

At the point of loading, this will use 29.158 GB of VRAM.

Setting a function in fish so you just have to type "qwen" in the terminal to launch you llama-server

Create a file called "/home/yourusername/.config/fish/functions/qwen.fish" and paste the following in there, obviously with the correct path.

function qwen
    bash -c "source /opt/intel/oneapi/setvars.sh && /opt/llama.cpp-sycl/bin/llama-server --model /path/to/your/model/Qwen3.8-27B-UD-Q6_K_XL.gguf --device SYCL0 --n-gpu-layers 999 --load-mode none --flash-attn on --jinja --reasoning-preserve --ctx-size 100000 --cache-type-k q8_0 --cache-type-v q8_0 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.00 --presence-penalty 0.0 --repeat-penalty 1.0 --spec-type draft-mtp --spec-draft-n-max 2 --port 9931"
end

Example 2: UD-Q4_K_XL, the highest quant you can run on a B70 that still allows you to use the full 262144 token context. Slightly reduced quality but a huge amount of context, ideal for tasks that don't require as much accuracy or intensive reasoning and instead work with a ton of data.

Compared to example 1, which should be prefered for most of your tasks, this drops the quant from Q6 to Q4, increases the cache from 100000 to 262144 tokens, reduces the reasoning for xhigh to medium and ups the draft tokens from 2 to 3. The cache quant is still kept at Q8 as dropping the model to Q4 gives less of a reduction in quality than dropping the quant of the cache.

Both of these are just examples from my experience, please do your own research and experimentation as it's both fun and rewarding to learn about this technology and adapt it to your own needs.

source /opt/intel/oneapi/setvars.sh
/opt/llama.cpp-sycl/bin/llama-server \
  --model /path/to/your/model/Qwen3.8-27B-UD-Q4_K_XL.gguf \
  --device SYCL0 \
  --n-gpu-layers 999 \
  --load-mode none \
  --flash-attn on \
  --jinja \
  --reasoning-preserve \
  --ctx-size 100000 \
  --cache-type-k q8_0 \
  --cache-type-v q8_0 \
  --temp 1.0 \
  --top-p 0.95 \
  --top-k 20 \
  --min-p 0.00 \
  --presence-penalty 0.0 \
  --repeat-penalty 1.0 \
  --chat-template-kwargs '{"reasoning_effort":"medium"}' \
  --spec-type draft-mtp \
  --spec-draft-n-max 3 \
  --port 9931

At the point of loading, this will use 29.031 GB of VRAM.

Setting a function in fish so you just have to type "qwenmaxctx" in the terminal to launch you llama-server

Create a file called "/home/yourusername/.config/fish/functions/qwenmaxctx.fish" and paste the following in there, obviously with the correct path.

function qwenmaxctx
    bash -c "source /opt/intel/oneapi/setvars.sh && /opt/llama.cpp-sycl/bin/llama-server --model /path/to/your/model/Qwen3.8-27B-UD-Q4_K_XL.gguf --device SYCL0 --n-gpu-layers 999 --load-mode none --flash-attn on --jinja --reasoning-preserve --ctx-size 262144 --cache-type-k q8_0 --cache-type-v q8_0 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.00 --presence-penalty 0.0 --repeat-penalty 1.0 --chat-template-kwargs '{\"reasoning_effort\":\"medium\"}' --spec-type draft-mtp --spec-draft-n-max 3 --port 9931"
end

---

License

PKGBUILD and packaging: MIT. llama.cpp: MIT. Intel oneAPI components are subject to Intel's license terms.