snap: add SYCL backend for Intel GPUs on amd64 by IsaacJT · Pull Request #322 · talhaHavadar/snap-llama.cpp

IsaacJT · 2026-09-24T09:23:42Z

The SYCL backend is compiled with the Intel oneAPI toolchain and targets Intel GPUs on amd64. Other architectures and GPUs are not supported with this backend.

This can be configured for use in the same way as the other backends.

Example of using this backend on an 11th Gen Intel CPU with Iris Xe graphics:

$ sudo snap run llama-cpp bench -hf empero-ai/Qwen3.8-4B-Distill-GGUF
load_backend: loaded CPU backend from /snap/llama-cpp/x14/bin/libggml-cpu-icelake.so
Build with Macros:
  GGML_SYCL_DNNL: yes
  GGML_SYCL_F16: no
  GGML_SYCL_FORCE_MMQ: no
  GGML_SYCL_GRAPH: yes
  GGML_SYCL_SUPPORT_LEVEL_ZERO_API: yes
  GGML_SYCL_SUPPORT_VMM: yes
Running with Environment Variables:
  GGML_SYCL_DEBUG: 0
  GGML_SYCL_DEV_DEBUG: 0
  GGML_SYCL_DEV2DEV_MEMCPY: 0 (SYCL API)
  GGML_SYCL_GET_MEM_API: 0 (Level Zero API)
  GGML_SYCL_ENABLE_DNN: 1
  GGML_SYCL_FA_ONEDNN: 1
  GGML_SYCL_FA_ONEDNN_MAX_KV: 0
  GGML_SYCL_ENABLE_MKL_FA: 1
  GGML_SYCL_MEMTRACE: 0
  GGML_SYCL_MEMTRACE_STEP: 64
  GGML_SYCL_ENABLE_FLASH_ATTN: 1
  GGML_SYCL_ENABLE_GRAPH: 0
  GGML_SYCL_ENABLE_OPT: 1
  GGML_SYCL_ENABLE_VMM: 1
  GGML_SYCL_ENABLE_FUSION: 1
  GGML_SYCL_ENABLE_ESIMD: 1
  GGML_SYCL_PRIORITIZE_DMMV: 0
  GGML_SYCL_USE_ASYNC_MEM_OP: 1
  GGML_SYCL_USE_LEVEL_ZERO_API: 1
  GGML_SYCL_USM_SYSTEM: 0
  GGML_SYCL_ENABLE_HOST_PINNED_MEM: 1
  GGML_SYCL_HOST_PINNED_MEM_2G: 0
Found 1 SYCL devices:
|  |                   |                                       |       |Max    |        |Max  |Global |                     |
|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
| 0| [level_zero:gpu:0]|                 Intel Iris Xe Graphics|   12.0|     96|     512|   32| 30947M|           1.14.37020|
SYCL Optimization Feature:
|ID|        Device Type|Reorder|
|--|-------------------|-------|
| 0| [level_zero:gpu:0]|      Y|
load_backend: loaded SYCL backend from /snap/llama-cpp/x14/../components/x14/sycl/lib/ggml/backends/libggml-sycl.so
| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| qwen35 4B Q4_K - Medium        |   2.58 GiB |     4.33 B | SYCL       |  -1 |           pp512 |        108.74 ± 0.36 |
| qwen35 4B Q4_K - Medium        |   2.58 GiB |     4.33 B | SYCL       |  -1 |           tg128 |          8.02 ± 0.03 |

build: b29c606 (1)

frenchwr · 2026-09-24T14:02:30Z

  echo
  echo "Backend: ${backend:-cpu}"
-  echo "  Configure persistently: sudo snap set llama-cpp backend={cpu,hip,cuda,vulkan,opencl,auto}"
+  echo "  Configure persistently: sudo snap set llama-cpp backend={cpu,hip,cuda,vulkan,opencl,cycl,auto}"


Typo: sycl

Thanks, fixed

frenchwr · 2026-09-24T14:10:17Z

+      - -DGGML_BUILD_TESTS=OFF
+      - -DGGML_BUILD_EXAMPLES=OFF
+      - -DCMAKE_VERBOSE_MAKEFILE=ON
+      - -DCMAKE_CXX_COMPILER=/opt/intel/oneapi/compiler/latest/bin/icpx


I'm curious if you tried building with the open source compiler from the archive:

build-packages: - on amd64: . . - dpclang-6 cmake-parameters: . - -DCMAKE_CXX_COMPILER=/usr/bin/dpclang++-6 - -DCMAKE_C_COMPILER=/usr/bin/dpclang-6

I'm guessing eventually you would see errors/conflicts because oneMKL, oneDNN, etc I assume were built with the closed source compiler from Intel, but I'm just curious if you tried at all.

You mean with the open source compiler but with the oneAPI libraries? I haven't tried that combination yet

Using dpclang++6 seems to override some environment settings or something for SYCL, CMake fails to configure:

2026-09-24 16:44:23.415 :: CMake Error at src/ggml-sycl/CMakeLists.txt:99 (target_link_libraries): 2026-09-24 16:44:23.415 :: Target "ggml-sycl" links to: 2026-09-24 16:44:23.415 :: 2026-09-24 16:44:23.415 :: IntelSYCL::SYCL_CXX 2026-09-24 16:44:23.415 :: 2026-09-24 16:44:23.415 :: but the target was not found. Possible reasons include: 2026-09-24 16:44:23.415 :: 2026-09-24 16:44:23.415 :: * There is a typo in the target name. 2026-09-24 16:44:23.416 :: * A find_package call is missing for an IMPORTED target. 2026-09-24 16:44:23.416 :: * An ALIAS target is missing. 2026-09-24 16:44:23.416 :: 2026-09-24 16:44:23.416 :: 2026-09-24 16:44:23.416 :: 2026-09-24 16:44:23.417 :: -- Generating done (0.0s) 2026-09-24 16:44:23.417 :: CMake Generate step failed. Build files cannot be regenerated correctly.

Might work with some more effort but not sure if it's worth it when we need to use the Intel Apt repo anyway

Right, the open source compiler is very close to feature parity with the closed source compiler (icp, icpx, which you get from Intel's repo). I don't think it would work, I was just curious if you had tried.

Might work with some more effort but not sure if it's worth it when we need to use the Intel Apt repo anyway

Yeah totally agree.

frenchwr · 2026-09-24T14:13:44Z

+    build-packages:
+      - on amd64:
+        - intel-oneapi-compiler-dpcpp-cpp-2026.1
+        - intel-oneapi-dnnl-devel-2026.0


Just a FYI, we also have oneDNN in the archive built with the open source oneAPI compiler (libdnnl-sycl3, libdnnl-sycl-dev). I'm not expecting it would work, but if in the future we figure out licensing issues blocking oneMKL from getting into the archive, it appears to me that everything else needed to build the backend is already in the archive, which is good news.

frenchwr · 2026-09-24T14:29:09Z

+      rm -f "${CRAFT_PART_INSTALL}"/usr/lib/"${CRAFT_ARCH_TRIPLET_BUILD_FOR}"/libOpenCL.so*
+
+      # Override the built-in .icd file as it contains a hard-coded path
+      printf '%s\n' 'libigdrcl.so' > \


In some snaps I've worked on I used a layout the overcome hard-coded paths, for example here in the audacity snap. I don't recall exactly where a path was hard-coded so not sure if it's exactly what you're fixing here but it is openCL related. Only mentioning it in case you prefer that approach instead, although maybe you don't want to add the layout since it's set even if you don't have the sycl component installed.

Unfortunately because we're using components layouts don't work (afaik)

The SYCL backend is compiled with the Intel oneAPI toolchain and targets Intel GPUs on amd64. Other architectures and GPUs are not supported with this backend. This can be configured for use in the same way as the other backends.

talhaHavadar · 2026-10-02T13:25:31Z

@frenchwr how does it look now?

frenchwr · 2026-10-02T14:53:51Z

@frenchwr how does it look now?

LGTM! I didn't test, though.

frenchwr reviewed Sep 24, 2026

View reviewed changes

snap: add SYCL backend for Intel GPUs on amd64

a01e707

The SYCL backend is compiled with the Intel oneAPI toolchain and targets Intel GPUs on amd64. Other architectures and GPUs are not supported with this backend. This can be configured for use in the same way as the other backends.

IsaacJT force-pushed the itrue/sycl branch from f97c531 to a01e707 Compare September 25, 2026 07:36

talhaHavadar merged commit d4a7ad9 into talhaHavadar:main Oct 6, 2026
3 of 5 checks passed

Navigation Menu

Repository navigation

snap: add SYCL backend for Intel GPUs on amd64 - #322

snap: add SYCL backend for Intel GPUs on amd64#322
talhaHavadar merged 1 commit into
talhaHavadar:mainfrom
IsaacJT:itrue/sycl

IsaacJT commented Sep 24, 2026 •

edited

Loading

Uh oh!

frenchwr Sep 24, 2026

Uh oh!

IsaacJT Sep 24, 2026

Uh oh!

frenchwr Sep 24, 2026

Uh oh!

IsaacJT Sep 24, 2026

Uh oh!

IsaacJT Sep 24, 2026

Uh oh!

frenchwr Sep 24, 2026

Uh oh!

frenchwr Sep 24, 2026

Uh oh!

frenchwr Sep 24, 2026

Uh oh!

frenchwr Sep 24, 2026

Uh oh!

IsaacJT Sep 24, 2026

Uh oh!

talhaHavadar commented Oct 2, 2026

Uh oh!

frenchwr commented Oct 2, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

IsaacJT commented Sep 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

talhaHavadar commented Oct 2, 2026

Uh oh!

frenchwr commented Oct 2, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

IsaacJT commented Sep 24, 2026 •

edited

Loading