b9949

ggml-org/llama.cppb9949Jul 10, 2026by github-actions[bot]same day as b9948

What changed since b9948

low risk

b9949 adds OpenCL cluster-parallel decode FlashAttention for Adreno GPUs; no other changes noted.

New features

  • OpenCL backend gains a cluster-parallel FlashAttention decode path targeting Adreno GPUs on Android and Windows.

Improvements

  • Faster FlashAttention decode on Adreno-based devices via OpenCL workgroup/cluster parallelism.

Upgrade notes

This is a single-PR increment over b9948, so no config or model changes are needed. Adreno users on Android or Windows should see decode speedups automatically when using the OpenCL backend; other backends are unaffected. The CUDA top-k/argsort memory optimization landed in b9948 and is unchanged here.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

Adds OpenCL cluster-parallel decode FlashAttention support for Adreno GPUs, optimizing performance on Android and Windows devices.

Key Highlights

  • OpenCL cluster-parallel decode FA for Adreno
  • Support for Adreno GPUs on Android and Windows

Release Notes (excerpt)

<details open> opencl: cluster-parallel decode FA for Adreno (#25473) </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