b9932

ggml-org/llama.cppb9932Jul 8, 2026by github-actions[bot]same day as b9931

What changed since b9931

low risk

Vulkan backend tuning: FA mask optimization disabled on GCN GPUs and re-enabled for head size 256.

Improvements

  • Vulkan flash-attention mask optimization is now disabled on GCN-class GPUs, which the release notes describe as a performance improvement for those cards.
  • Mask optimization is re-enabled for attention head size 256, restoring the faster path for models using that head dimension.
  • Only the Vulkan backend is touched; OpenCL MoE ragged-tile work from the previous build carries over unchanged.

Upgrade notes

This is a Vulkan-only tuning build with no config or API changes, so a straight binary swap is sufficient. GCN users should see better flash-attention throughput; no environment variables or flags need adjusting. Users on other backends gain nothing from this bump and can stay on the prior build if desired.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Vulkan backend performance tuning for GCN GPUs and attention masking optimization.

Key Highlights

  • Disabled FA mask_opt on GCN to improve performance
  • Re-enabled mask optimization for attention head size 256

Release Notes (excerpt)

<details open> vulkan: disable FA mask_opt on GCN to improve performance (#24362) * vulkan: disable FA mask_opt on GCN to improve performance * reenable mask opt over attention head size 256 </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)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories