b9906

ggml-org/llama.cppb9906Jul 8, 2026by github-actions[bot]same day as b9905

What changed since b9905

low risk

b9906 is a build-only fix that restores INFINITY/NaN handling in ROCm/clang 22 HIP builds by adding -fno-finite-math-only.

Improvements

  • ggml-hip now passes -fno-finite-math-only alongside -ffast-math, keeping fast-math optimizations while re-enabling IEEE infinity and NaN semantics.
  • Prebuilt binaries were regenerated for all supported platforms at the new revision.

Bug fixes

  • Fixes ROCm/clang 22 compilation failures where -ffast-math implied -ffinite-math-only and triggered -Wnan-infinity-disabled errors under -Werror in CI.

Upgrade notes

No API or runtime behavior changes; only HIP/ROCm compilation is affected. Rebuild from source if you compile with ROCm and clang 22+, or if you previously worked around the warning by dropping -ffast-math. Users of prebuilt non-HIP binaries can upgrade with no action needed.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Corrects a compilation error where `-ffast-math` was inadvertently disabling Infinity and NaN handling under ROCm 22, which caused build failures in CI.

Key Highlights

  • Add `-fno-finite-math-only` flag to ggml-hip to re-enable Infinity/NaN handling
  • Fixes CI build errors related to `-Wnan-infinity-disabled`

New Features

  • Compiler flag configuration fix

Release Notes (excerpt)

<details open> ggml-hip : add -fno-finite-math-only alongside -ffast-math (#25373) -ffast-math implies -ffinite-math-only under ROCm/clang 22, which disables INFINITY/NaN and triggers -Wnan-infinity-disabled (errors under -Werror in CI). Re-enable infinity handling without dropping the rest of fast-math. Fixes #25361 </details> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories