b10034

ggml-org/llama.cppb10034Jul 15, 2026by github-actions[bot]same day as b10032

What changed since b10032

low risk

b10034 is a targeted OpenCL fix for corrupted MoE output on Adreno GPUs, with no changes to b10032's CUDA lightning indexer work.

Improvements

  • Adreno A6x and unrecognized Adreno devices are now excluded from MoE weight repacking, avoiding the miscompiled path entirely.
  • Kernel selection for OpenCL MoE is now gated by GPU architecture rather than applied uniformly across Adreno families.

Bug fixes

  • MoE models no longer produce garbage output on Adreno A7x GPUs where certain compilers miscompiled the repack kernels and corrupted weights.
  • Repack kernels are skipped on A6x and unknown Adreno devices, closing a second route to weight corruption.

Upgrade notes

This is a drop-in rebuild with no config or API changes; only OpenCL users on Adreno hardware are affected. If you run MoE models via the OpenCL backend on A7x/A6x devices, upgrade to restore correct output. The CUDA lightning indexer feature from b10032 carries over unchanged.

Analyzed Nov 2, 2026

Upgrading?

Release notes summary

Fixes a bug in OpenCL MoE (Mixture of Experts) kernels for specific Adreno GPU architectures that were causing corrupted output.

Key Highlights

  • Excludes Adreno A7x from using Adreno MoE kernels
  • Excludes A6x and unknown Adreno from MoE weights repack
  • Fixes corrupted output on specific Adreno devices

New Features

  • Bug fix for OpenCL MoE kernels

Release Notes (excerpt)

<details open> opencl: exclude some moe kernels on Adreno a7x (#25698) * opencl: exclude Adreno A7x from using Adreno MoE kernels Some compilers for A7x devices miscompile the repack kernels, corrupting the weights and causing MoE models to generate garbage output * opencl: exclude A6x and unknown Adreno from MoE weights repack </details> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories