b9968

ggml-org/llama.cppb9968Jul 12, 2026by github-actions[bot]same day as b9967

What changed since b9967

low risk

b9968 adds Adreno-specific OpenCL int8 DP4 dense matmul and MoE prefill kernels, with no API changes from b9967.

New features

  • OpenCL backend now supports int8 DP4 dense matrix multiplication optimized for Adreno GPUs.
  • Added a MoE (Mixture of Experts) prefill GEMM path for Adreno GPUs via OpenCL.

Improvements

  • OpenCL implementation was refactored to accommodate the new Adreno kernels.
  • Faster int8 dense and MoE prefill inference on Qualcomm Adreno hardware.

Upgrade notes

No configuration or API changes are required; the new kernels are selected automatically on supported Adreno/OpenCL devices. Users on other backends (CUDA, Metal, Vulkan, CPU) should see no behavioral difference from b9967.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

Added optimized int8 dense matrix multiplication and MoE (Mixture of Experts) prefill operations specifically for Adreno GPUs via the OpenCL backend.

Key Highlights

  • Added int8 DP4 dense matrix multiplication support for Adreno GPUs
  • Added MoE prefill optimization for Adreno GPUs
  • Refactored OpenCL implementation

Release Notes (excerpt)

<details open> opencl: add int8 dp4 dense and MoE prefill optimization for Adreno GPUs (#25537) * opencl: add int8 dp4 dense and moe GEMM * opencl: refactor --------- Co-authored-by: Li He <lih@qti.qualcomm.com> </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)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories