b10156

ggml-org/llama.cppb10156Jul 28, 2026by github-actions[bot]same day as b10155

What changed since b10155

low risk

b10156 only disables the -ffast-math compiler flag for the HIP backend to improve numerical precision.

Improvements

  • HIP (Radeon) builds no longer compile with -ffast-math, improving floating-point accuracy on AMD GPUs.
  • The change is confined to the HIP build configuration, leaving other backends and prebuilt binaries unchanged.

Bug fixes

  • Addresses numerical precision drift in HIP computations that could arise from aggressive fast-math optimizations.

Upgrade notes

No API, model, or configuration changes are involved; the delta is a single build-flag adjustment for HIP. Users on ROCm/HIP should rebuild to pick up the more precise binaries, and may notice a small performance difference versus b10155. Non-HIP users can upgrade with no expected impact.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Disables the -ffast-math compiler flag for HIP (Radeon) support to improve numerical precision in calculations.

Key Highlights

  • Disabled -ffast-math for HIP backend
  • Cross-platform binaries maintained

New Features

  • HIP numerical precision fix

Release Notes (excerpt)

<details open> Disable -ffast-math on HIP (#25495) </details> **Website:** - <https://llama.app> **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