Skip to content

[SYCL] Support DSv4 OPs: LIGHTNING_INDEXER,DSV4_HC_COMB,DSV4_HC_POST,DSV4_HC_PRE - #26568

Merged
ggerganov merged 4 commits into
ggml-org:masterfrom
arthw:add_ops_dsv4
Aug 7, 2026
Merged

[SYCL] Support DSv4 OPs: LIGHTNING_INDEXER,DSV4_HC_COMB,DSV4_HC_POST,DSV4_HC_PRE#26568
ggerganov merged 4 commits into
ggml-org:masterfrom
arthw:add_ops_dsv4

Conversation

@arthw

@arthw arthw commented Aug 4, 2026

Copy link
Copy Markdown
Contributor

Overview

Fix issue: #26549

Additional information

Support DSv4 OPs:

  • LIGHTNING_INDEXER
  • DSV4_HC_COMB
  • DSV4_HC_POST
  • DSV4_HC_PRE

All related UT cases are passed.
Thank @mobilinkd to provide the patches.

To support DeepSeek-V4-Flash, need the solution in PR: #26515.

Requirements

@arthw
arthw requested a review from a team as a code owner August 4, 2026 10:46
@github-actions github-actions Bot added documentation Improvements or additions to documentation ggml changes relating to the ggml tensor library for machine learning SYCL https://en.wikipedia.org/wiki/SYCL - GPU programming language labels Aug 4, 2026
@jaroslavstreit

jaroslavstreit commented Aug 4, 2026

Copy link
Copy Markdown

Just bought Intel Arc Pro B65 and have problems with running DeepSeek V4 0731 - DSV4, so this could solve it, thank you, will let you know the results then. Without it, on Vulkan, got 7-9 t/s, 5070Ti got around 20 - because CUDA supports it.

EDIT: tested it, +11% on generating, -12% on prefill, totally worth it, thanks! DeepSeek-V4-Flash UD-IQ1_S — 76,9 GiB, context 16 384, --n-cpu-moe 43 (all 43 in RAM), KV cache f16, Ryzen 9950x, 96GB RAM 6000MHz, Arc B65 32GB.

@arthw

arthw commented Aug 5, 2026

Copy link
Copy Markdown
Contributor Author

@jaroslavstreit
Got it!
How many B65 GPUs in your case?

Thank you for your sharing!

@arthw arthw added the merge ready A maintainer can use this label to indicate that they consider the changes final and ready to merge. label Aug 5, 2026
@jaroslavstreit

Copy link
Copy Markdown

@jaroslavstreit
Got it!
How many B65 GPUs in your case?

Thank you for your sharing!

Just one.

@arthw

arthw commented Aug 5, 2026

Copy link
Copy Markdown
Contributor Author

DeepSeek-V4-Flash UD-IQ1_S is big: 76.9 GiB.
How to run it on one GPU: 32GB?
Do you load a part to GPU?

If yes, the performance will be impacted more.

Thank you!

@jaroslavstreit

Copy link
Copy Markdown

Rechecked it now with 24GB on VRAM, working stable. Now doing more tests and comparison with Vulkan. Plus then test with 200 000 ctx. Had error device_not_found when VRAM overflowed, but this is not code bug, it's wrong settings problem.

@jaroslavstreit

jaroslavstreit commented Aug 5, 2026

Copy link
Copy Markdown
image

I have Ryzen 9 9950X, Crucial PRO 96 GB DDR5-6000 · ASUS ROG CROSSHAIR 2006 X870E.

@ggerganov
ggerganov merged commit 31558db into ggml-org:master Aug 7, 2026
22 of 30 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

documentation Improvements or additions to documentation ggml changes relating to the ggml tensor library for machine learning merge ready A maintainer can use this label to indicate that they consider the changes final and ready to merge. SYCL https://en.wikipedia.org/wiki/SYCL - GPU programming language

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants