ci : update the oneAPI toolkit to 2026.1

[Asahi-Prv]

Overview

Updates the CI build environment and the release package build from oneAPI 2025.3.3 (Deep Learning Essentials) to oneAPI 2026.1 (the unified oneAPI Toolkit installer). Since oneDNN was removed from the Deep Learning Essentials package in oneAPI 2026.0, this keeps oneDNN support available through the unified Toolkit and updates the corresponding SYCL documentation.

Additional information

Changes

• Update the CI build environment from oneAPI 2025.3.3 (Deep Learning Essentials) to oneAPI 2026.1 (the unified oneAPI Toolkit installer).

• Keep oneDNN support available with the unified oneAPI Toolkit. oneDNN was removed from the Deep Learning Essentials package in oneAPI 2026.0.

• The Windows install script (scripts/install-oneapi.bat) works as-is because the component IDs are unchanged in oneAPI 2026.1.

• Align the release package build (.github/workflows/release.yml):

  • Level Zero SDK 1.28.2 -> 1.33.1 (the libze1/libze-dev Debian package names).
  • The Windows runtime DLL list for the 2026.1 versions (sycl9.dll, .6/.3 MKL libraries).

• Update docs/backend/SYCL.md:

  • Update the installation instructions.
  • Add oneAPI 2026.1 to the verified release table.
  • Update the news section.

Verification

Using the same code (b10899) on an Intel Arc B570 with Qwen3.8-9B Q4_K_M and -ngl 99:

• Prompt processing (pp512): 1331 t/s with oneAPI 2026.1 vs 434 t/s with the 2025.3-based release build (~3.1x).
• Token generation (tg128): 50.1 t/s with oneAPI 2026.1 vs 45.3-48.0 t/s with the 2025.3-based release build.
• test-backend-ops: correct results verified with the oneAPI 2026.1 build.

Requirements

• I have read and agree to the contributing guidelines

• AI usage disclosure: YES - GLM-5.3 Flash via OpenCode was used for analysis, implementation, routine edits, and debugging assistance. The author identified the issue and provided the overall direction for the fix.

[CISC]

Provide link to successful release job please.

[Asahi-Prv]

@CISC I ran the release workflow with the updated config. All jobs are green (28/28), including windows-sycl and ubuntu-24-sycl (fp32/fp16) with oneAPI 2026.1 + Level Zero SDK 1.33.1:
https://github.com/Asahi-Prv/llama.cpp/actions/runs/35826107843

[arthw]

The oneapi used in dockerfile need to be updated in same time:

update .devops/intel.Dockerfile with intel/oneapi-toolkit:2026.1.1-devel-ubuntu24.04

Thank you!

[Asahi-Prv]

I forgot left comment. I fixed dockerfile. Thank you!

[WizardlyBump17]

I would like to speak against the upgrade on the intel.Dockerfile. It is known that the base oneAPI 2026.1 image causes performance decrease: #25973

[NickM-27]

@WizardlyBump17 The change here doesn't just bump the version, it also changes to a different image, one which contains oneDNN (unlike the updated version of the original image).

I tested on B70 and I see no performance difference:

current master

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 Arc Pro B70 Graphics|   20.2|    256|    1024|   32| 34242M|         1.15.38308+1|
SYCL Optimization Feature:
|ID|        Device Type|Reorder|
|--|-------------------|-------|
| 0| [level_zero:gpu:0]|      Y|
load_backend: loaded SYCL backend from /app/libggml-sycl.so
load_backend: loaded CPU backend from /app/libggml-cpu-haswell.so
| model                          |       size |     params | backend    | ngl | n_ubatch |  fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | --: | --------------: | -------------------: |
| qwen35 27B Q6_K                |  20.46 GiB |    27.32 B | SYCL       |  -1 |     1024 |   1 |          pp2048 |       1212.47 ± 9.64 |
| qwen35 27B Q6_K                |  20.46 GiB |    27.32 B | SYCL       |  -1 |     1024 |   1 |           tg256 |         21.32 ± 0.05 |
| qwen35 27B Q6_K                |  20.46 GiB |    27.32 B | SYCL       |  -1 |     1024 |   1 |  pp2048 @ d8192 |       1029.24 ± 2.03 |
| qwen35 27B Q6_K                |  20.46 GiB |    27.32 B | SYCL       |  -1 |     1024 |   1 |   tg256 @ d8192 |         19.60 ± 0.03 |
| qwen35 27B Q6_K                |  20.46 GiB |    27.32 B | SYCL       |  -1 |     1024 |   1 | pp2048 @ d16384 |        968.59 ± 1.90 |
| qwen35 27B Q6_K                |  20.46 GiB |    27.32 B | SYCL       |  -1 |     1024 |   1 |  tg256 @ d16384 |         18.15 ± 0.02 |

this PR

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 Arc Pro B70 Graphics|   20.2|    256|    1024|   32| 34242M|         1.15.38308+1|
SYCL Optimization Feature:
|ID|        Device Type|Reorder|
|--|-------------------|-------|
| 0| [level_zero:gpu:0]|      Y|
load_backend: loaded SYCL backend from /app/libggml-sycl.so
load_backend: loaded CPU backend from /app/libggml-cpu-haswell.so
| model                          |       size |     params | backend    | ngl | n_ubatch |  fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | --: | --------------: | -------------------: |
| qwen35 27B Q6_K                |  20.46 GiB |    27.32 B | SYCL       |  -1 |     1024 |   1 |          pp2048 |      1215.89 ± 10.86 |
| qwen35 27B Q6_K                |  20.46 GiB |    27.32 B | SYCL       |  -1 |     1024 |   1 |           tg256 |         21.36 ± 0.04 |
| qwen35 27B Q6_K                |  20.46 GiB |    27.32 B | SYCL       |  -1 |     1024 |   1 |  pp2048 @ d8192 |       1034.93 ± 0.35 |
| qwen35 27B Q6_K                |  20.46 GiB |    27.32 B | SYCL       |  -1 |     1024 |   1 |   tg256 @ d8192 |         19.66 ± 0.03 |
| qwen35 27B Q6_K                |  20.46 GiB |    27.32 B | SYCL       |  -1 |     1024 |   1 | pp2048 @ d16384 |        976.01 ± 0.95 |
| qwen35 27B Q6_K                |  20.46 GiB |    27.32 B | SYCL       |  -1 |     1024 |   1 |  tg256 @ d16384 |         18.23 ± 0.01 |

[Asahi-Prv]

@WizardlyBump17 Yeah, I knew this issue.
Reason is simple. Separation oneDNN Package by oneAPI Toolkits.
oneDNN is bundled my changed dockerfile.
Thank you!

[WizardlyBump17]

@WizardlyBump17 Yeah, I knew this issue. Reason is simple. Separation oneDNN Package by oneAPI Toolkits. oneDNN is bundled my changed dockerfile. Thank you!

I didnt realize that, sorry. But it would be cooler if it still used the deep-learning-essentials image as it is smaller by 3GB:

[NeoZhangJianyu]

It's good job!

Thank you!

[arthw]

@WizardlyBump17
deep-learning-essentials (DLE) has removed the oneDNN package from it.
oneDNN is mandatory package of SYCL backend.
It's simple way to install from oneAPI toolkit package.

I will consider to install DLE+oneDNN in the future.

Thank you!