b10408

ggml-org/llama.cppb10408Aug 13, 2026by github-actions[bot]same day as b10405

What changed since b10405

medium risk

Adds SYCL ESIMD support for Intel GPUs while fixing HIP math optimizations.

New features

  • SYCL ESIMD support for Q3_K, Q4_K, and Q6_K kernels
  • DMMV ESIMD Q3_K kernel

Improvements

  • Refactor ESIMD kernels to share common code
  • Move control of ESIMD from compile to runtime
  • Use ESIMD by default when available

Bug fixes

  • Fix possible error when using ESIMD by default

Upgrade notes

Enable SYCL ESIMD via CMake flag or rely on the new default behavior for Intel GPUs. No configuration changes are required for HIP builds.

Analyzed Oct 15, 2026

Upgrading?

Release notes summary

Adds support for SYCL ESIMD (Exponential-SIMD) kernels for multiple quantization types, improving performance on Intel GPUs.

Key Highlights

  • Add DMMV ESIMD Q3_K kernel (#26251)
  • Refactor ESIMD kernels to share common code
  • Move control of ESIMD from compile to runtime

Release Notes (excerpt)

<details open> sycl : Add DMMV ESIMD Q3_K kernel (#26251) * Add DMMV Q4_K and Q6_K ESIMD kernels Configure cmake build with -DGGML_SYCL_ESIMD=ON to enable. Signed-off-by: Todd Malsbary <todd.malsbary@intel.com> * Refactor ESIMD kernels to share common code Signed-off-by: Todd Malsbary <todd.malsbary@intel.com> * Move control of ESIMD from compile to runtime Signed-off-by: Todd Malsbary <todd.malsbary@intel.com> * Use ESIMD by default when available…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories