b10036

ggml-org/llama.cppb10036Jul 16, 2026by github-actions[bot]same day as b10035

What changed since b10035

low risk

b10036 adds an OpenCL workaround disabling FA and MoE weight repack on Adreno 850 GPUs and fixes DirectX compiler version parsing.

New features

  • OpenCL backend now detects Adreno 850 GPUs and skips flash-attention and MoE weight repacking to avoid compiler failures.

Improvements

  • DirectX compiler version parsing was corrected and the surrounding detection code cleaned up.
  • OpenCL kernel selection is now GPU-aware, avoiding known-bad code paths on specific Adreno hardware.

Bug fixes

  • Resolved compiler-related errors/crashes on Adreno 850 by disabling the offending FA and MoE repack kernels.
  • Fixed incorrect parsing of DirectX compiler version strings.

Upgrade notes

No config or model changes are required; this is a drop-in update from b10035. Adreno 850 users should expect flash-attention and MoE weight repack paths to be skipped, which may reduce performance on that GPU but restores stability. Other backends (CUDA, Vulkan, ROCm, CPU) are unaffected.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Disables specific OpenCL kernels (FA and MoE) on Adreno 850 to prevent compiler-related errors, and fixes a parsing bug for DirectX compiler versions.

Key Highlights

  • Disables FA and MoE weights repack for Adreno 850
  • Fixes DX compiler version parsing and cleanup

Release Notes (excerpt)

<details open> opencl: disable FA and MoE weights repack to work around compiler issues for Adreno 850 GPU (#25745) * opencl: workaround for A850 compiler compat * opencl: fix DX compiler version parsing and cleanup --------- 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)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories