b10539

ggml-org/llama.cppb10539Aug 21, 2026by github-actions[bot]same day as b10538

What changed since b10538

low risk

b10539 adds a Vulkan Flash Attention MMQ fix using fp32 for Q quantization to prevent overflow from denormalized values.

Improvements

  • Vulkan FA MMQ kernel now performs Q quantization math in fp32 instead of relying on lower-precision values.
  • Guards against overflow when the quantization divisor is denormalized, improving numerical robustness on Vulkan.

Bug fixes

  • Fixed a potential overflow in the Vulkan Flash Attention MMQ kernel caused by denormalized Q values and reciprocal overflow.

Upgrade notes

This is a single-commit follow-up to b10538 with no API, config, or model-format changes. Vulkan users running Flash Attention benefit from the fix; other backends are unaffected. Simply rebuild or pull the new binaries.

Analyzed Oct 26, 2026

Upgrading?

Release notes summary

Fixes a potential overflow in the Vulkan Flash Attention (FA) MMQ kernel where Q quantization calculations used denormalized floats.

Key Highlights

  • Vulkan FA MMQ now uses fp32 for Q quantization calculations
  • Prevents overflow from denormalized Q values

New Features

  • Vulkan kernel stability fix

Release Notes (excerpt)

<details open> vulkan: FA MMQ should use fp32 for Q quantization calculations (#27413) Codex found that qd could be a denorm and 1/qd would overflow. </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42034841> **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)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories