b10256

ggml-org/llama.cppb10256Aug 4, 2026by github-actions[bot]same day as b10255

What changed since b10255

low risk

SYCL non-contiguous concat kernel is now parallelized, giving ~9.4% faster prompt processing on SYCL backends.

New features

  • SYCL non-contiguous concat kernel now launches a multi-lane work-group instead of a single lane, enabling parallel execution.

Improvements

  • Prompt processing throughput on SYCL improved roughly 9.4% in llama-bench (pp2048 920 -> 1006 t/s on Arc Pro B70).
  • Block width for the concat kernel is capped at ne0 to avoid oversizing work-groups on small tensors.
  • SYCL_CONCAT_BLOCK_SIZE is now defined in ggml/src/ggml-sycl/presets.hpp for easier tuning.

Bug fixes

  • Fixed the block size cap so the parallel concat launch geometry stays valid for narrow tensors.

Upgrade notes

This is a SYCL-only performance change with no API or config impact; non-SYCL builds are unaffected. An env-tunable block size was proposed but reverted, so tuning requires editing presets.hpp and rebuilding. No migration steps are needed.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Optimizes SYCL performance by parallelizing the non-contiguous concatenation kernel.

Key Highlights

  • Parallelized non-contiguous concat kernel
  • Performance improvement (~9.4%)
  • Fixed block size cap

Release Notes (excerpt)

<details open> sycl: parallelize the non-contiguous concat kernel (#25852) * sycl: parallelize the non-contiguous concat kernel Launch geometry only: the non-contiguous concat kernel launched a single-lane work-group (1, 1, 1), now it will launch a (1, 1, SYCL_CONCAT_BLOCK_SIZE) one. SYCL_CONCAT_BLOCK_SIZE is defined in `ggml/src/ggml-sycl/presets.hpp`. llama-bench (Arc Pro B70, Qwen3.6-27B-UD-Q4_K_XL, -fa on, q8_0 KV), on top of upstream master: pp2048 920 -> 1006 t/s (+9.4%)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories