b9651

ggml-org/llama.cppb9651Jun 15, 2026by github-actions[bot]same day as b9650

What changed since b9650

low risk

SYCL backend gets a K-quant DMMV optimization using native subgroup size, on top of the prior softmax fix.

New features

  • SYCL DMMV kernel for K-quant models now uses the device's native subgroup size for better efficiency.

Improvements

  • K-quant matrix-vector multiplication on SYCL devices should run faster due to subgroup-size-aware dispatch.
  • Builds on the previous release's corrected SYCL soft_max_f32 max reduction, keeping SYCL numerics consistent.

Upgrade notes

This is a routine incremental bump with no config or API changes; SYCL users running K-quant models are the primary beneficiaries. Non-SYCL backends are unaffected, so upgrading is safe for all other platforms.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

This update optimizes the SYCL backend for K-quants by utilizing the native subgroup size for the DMMV operation to improve efficiency.

Key Highlights

  • SYCL: Uses native subgroup size for K-quant DMMV
  • Optimized performance for K-quants on SYCL devices

Release Notes (excerpt)

<details open> SYCL: use native subgroup size for K-quant DMMV (#21700) </details> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - Ubuntu x64 (CPU) - Ubuntu arm64 (CPU) - Ubuntu s390x (CPU) - Ubuntu x64 (Vulkan) - Ubuntu arm64 (Vulkan) - Ubuntu x64 (ROCm 7.2) - Ubuntu x64 (OpenVINO) - Ubuntu x64 (SYCL FP32) - Ubuntu x64 (SYCL FP16) **Android:** - Android arm64 (CPU) **Windows:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories