b9776

ggml-org/llama.cppb9776Jun 24, 2026by github-actions[bot]same day as b9775

What changed since b9775

low risk

Fixes Vulkan Fully Attention overflow by applying bias before softmax.

Bug fixes

  • Fixes Vulkan floating-point overflow in Fully Attention (FA) by applying bias before softmax.

Upgrade notes

No configuration changes are required for this patch release. The change is internal to the Vulkan backend.

Analyzed Oct 17, 2026

Upgrading?

Release notes summary

Fixes a Vulkan floating-point overflow issue in Fully Attention (FA) by applying bias before the softmax operation. Includes standard binary builds for macOS, Linux, Windows, Android, and openEuler.

Key Highlights

  • Vulkan FA bias overflow fix
  • macOS and iOS builds (arm64, x64)
  • Linux builds (ROCm, OpenVINO, SYCL)
  • Windows builds (CUDA 12/13, HIP, Vulkan)
  • Android ARM64 builds

New Features

  • Vulkan bias before softmax to prevent overflow

Release Notes (excerpt)

<details open> vulkan: Apply bias before softmax in FA, to avoid overflow (#24909) </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) - Ubuntu x64 (OpenVINO) - Ubuntu x64 (SYCL FP32) - Ubuntu x64 (SYCL FP16) **Android:** - Android arm64 (CPU)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories