b10405

ggml-org/llama.cppb10405Aug 13, 2026by github-actions[bot]same day as b10400

What changed since b10400

medium risk

Fixes ARM builds and removes unsafe math optimizations for HIP compliance.

Improvements

  • HIP builds now use IEEE-conformant floating-point reductions.

Bug fixes

  • Fixed ARM builds to compile successfully.
  • Fixed unused variable warnings in the ggml library.
  • Fixed greedy argmax divergence on RDNA3.5 HIP builds.

Upgrade notes

No configuration changes are required. Update to the new binary to resolve argmax divergence on RDNA3.5 hardware.

Analyzed Oct 15, 2026

Upgrading?

Release notes summary

Removes unsafe math optimizations from HIP builds to ensure IEEE-conformant behavior and prevent greedy argmax divergence on RDNA3.5.

Key Highlights

  • Remove -funsafe-math-optimizations to fix greedy argmax divergence (#26696)
  • Ensures IEEE-conformant FP reductions

Release Notes (excerpt)

<details open> ggml-hip : remove -funsafe-math-optimizations (#26696) It enables -fassociative-math, which reassociates FP reductions and can flip greedy argmax on RDNA3.5 (e.g. MTP speculative decode diverging from the non-speculative baseline). Drop it so HIP builds are IEEE-conformant. Co-authored-by: Jim Wu <ywu@xilinx.com> </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories