b9881

ggml-org/llama.cppb9881Jul 6, 2026by github-actions[bot]1 days after b9878

What changed since b9878

low risk

b9881 enables -ffast-math for HIP builds, speeding up floating-point math on AMD GPUs.

Improvements

  • HIP (AMD GPU) builds now compile with -ffast-math, improving floating-point throughput on AMD hardware.
  • The change is scoped to the ggml-hip backend build configuration, leaving other backends untouched.

Upgrade notes

This is a build-flag change for HIP only; no config or API changes are required. Because -ffast-math relaxes IEEE floating-point semantics, AMD GPU users who depend on bit-exact reproducibility should validate outputs after upgrading. Users on CPU, CUDA, Vulkan, or Metal builds are unaffected.

Analyzed Oct 28, 2026

Upgrading?

Release notes summary

This release enables the `-ffast-math` compiler flag for HIP (AMD GPU) builds, which optimizes floating-point arithmetic for improved performance on AMD hardware.

Key Highlights

  • Enable -ffast-math for HIP builds
  • Optimize floating-point performance for AMD GPUs

Release Notes (excerpt)

<details open> ggml-hip: enable -ffast-math for HIP builds (#23862) </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) **Windows:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories