SYCL: Add GGML_OP_MEAN operator support

[yael-works]

This PR introduces support for GGML_OP_MEAN in the SYCL backend.

Changes:

• Added ggml_sycl_op_mean() kernel and ggml_sycl_mean() wrapper
• Registered GGML_OP_MEAN in ggml_sycl_compute_forward()
• Updated ggml_backend_sycl_device_supports_op() to report support
• Added proper row-wise mean implementation with parallel division

Verification:

• Build passes successfully with CMake + make
• ./build/bin/test-backend-ops support -o MEAN -b sycl shows
• ./build/bin/test-backend-ops test -o MEAN -b sycl passes with OK

[yael-works]

Status update:
All major builds (Linux, Windows, SYCL, CUDA, macOS-x64) passed successfully.

RISCV job failed due to missing runner:

The job was not acquired by Runner of type self-hosted even after multiple attempts
Internal server error. Correlation ID: e44cad51-7cf6-459f-9d19-6597095ade25

This appears unrelated to the code changes.

@ggerganov @ggerganov/ggml-maintainers –
Please review when convenient. This PR is ready for review & merge.

[yael-works]

I checked the macOS-arm64 logs – the only failing test is test-barrier (timeout).
This seems unrelated to the SYCL/MEAN implementation.
All other tests (Linux, Windows, SYCL, CUDA, etc.) pass successfully.
Happy to address anything if you think changes are needed.

[yael-works]

Gentle ping @NeoZhangJianyu – just checking if you had a chance to review.
Happy to adjust anything if needed. Thanks!

[yael-works]

Hi @ggerganov @NeoZhangJianyu 👋
Just following up on this PR (#16009) — it’s been a month since submission.
All CI checks are still green, and the change is limited to adding GGML_OP_MEAN support for SYCL.

Please let me know if there’s anything I can adjust or clarify to help move it forward.
Thanks again for your time and great work maintaining this project!

[yael-works]

Hi @sigbjørnSkjæret
I noticed you recently commented on related PRs. My PR #16009 adds GGML_OP_MEAN support for SYCL and has been passing all CI checks for a month.

Could you please take a look or point me to the right maintainer to approve it?
I’d be happy to make any adjustments if needed.
Thanks so much for your help!

[sigbj]

Looks like this is a serious email, but email-search generated through ArtInt. It is not me. I utilise only bourneshell through vimrc forwarded to the ***@***.*** The guy You are searching maybe lives just outside Oslo. Take a look at the attached photo: ons. 15. okt. 2025, 12:03 skrev yael-works ***@***.***>:

…

*yael-works* left a comment (ggml-org/llama.cpp#16009) <#16009 (comment)> Hi @sigbj <https://github.com/sigbj>ørnSkjæret I noticed you recently commented on related PRs. My PR #16009 <#16009> adds GGML_OP_MEAN support for SYCL and has been passing all CI checks for a month. Could you please take a look or point me to the right maintainer to approve it? I’d be happy to make any adjustments if needed. Thanks so much for your help! — Reply to this email directly, view it on GitHub <#16009 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AVBTEZZ3QVEISNSXUYBBWOT3XYLX7AVCNFSM6AAAAACGRMCADGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTIMBVGU4TQMRSHE> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[yael-works]

Hi @CISC

Just wanted to follow up on PR #16009. All CI checks are green and the change is limited to adding GGML_OP_MEAN support for SYCL.

Would you be able to review when convenient? Happy to make any adjustments if needed. Thanks so much for your time!

[CISC]

@qnixsynapse Can you review/test this, looks like @NeoZhangJianyu is busy?

[qnixsynapse]

@CISC I no longer maintain SYCL and moved away from using Intel GPU. However, the code looks okay to me.

[CISC]

@ggerganov Looks like we no longer have any maintainers for the SYCL backend, unless @NeoZhangJianyu reappears...

[yael-works]

Hi @ggerganov

All CI checks are green (some still queued), and both @CISC and @qnixsynapse have reviewed the changes.
There are currently no active SYCL maintainers, so the PR can’t be approved automatically.

Could you please take a look and approve or merge when convenient?
This adds GGML_OP_MEAN support for the SYCL backend — tested and verified across platforms.

Thanks a lot for your time and great work maintaining the project!

[NeoZhangJianyu]

It's good work!

Thank you!