b9985

ggml-org/llama.cppb9985Jul 13, 2026by github-actions[bot]same day as b9984

What changed since b9984

low risk

b9985 adds an opt-in fused top-k MoE path for the SYCL backend, building on b9984's SYCL Q2_K reorder support.

New features

  • SYCL backend gains a fused top-k Mixture-of-Experts (MoE) implementation for faster expert routing on Intel hardware.
  • New GGML_SYCL_ENABLE_FUSION environment variable toggles the SYCL fusion path, and its state is logged at startup like other env vars.

Improvements

  • Fusion dispatch was placed in the top-k MoE code path rather than the generic kernel, keeping the change scoped to SYCL.
  • Release artifacts continue to cover macOS, Linux, Windows and Android across CUDA 12/13, ROCm, Vulkan, OpenVINO and SYCL builds.

Upgrade notes

This is a single-commit increment over b9984, so no config or model migration is required. SYCL users who want the new MoE speedup should set GGML_SYCL_ENABLE_FUSION and confirm the startup log shows it enabled; other backends are unaffected. Note that KleidiAI-enabled macOS Apple Silicon builds remain disabled, as in the previous release.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

This release introduces fused top-k Mixture-of-Experts (MoE) support for the SYCL backend, optimizing performance for specific hardware configurations. It also includes various binary builds across different platforms (macOS, Linux, Windows, Android) and architectures.

Key Highlights

  • Added fused top-k MoE support for SYCL backend
  • Introduced GGML_SYCL_ENABLE_FUSION environment variable
  • Disabled KleidiAI builds for macOS Apple Silicon
  • Expanded support for CUDA 12, CUDA 13, ROCm, Vulkan, OpenVINO, and SYCL backends

Release Notes (excerpt)

<details open> sycl: add fused top-k MoE (#25217) * sycl: add fused top-k MoE * sycl: address review: GGML_SYCL_ENABLE_FUSION env, move fusion dispatch to topk-moe * sycl: print GGML_SYCL_ENABLE_FUSION at startup like other env vars Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> </details> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories