b10310

ggml-org/llama.cppb10310Aug 7, 2026by github-actions[bot]same day as b10308

What changed since b10308

low risk

ARM64 CPU detection is hardened with HWCAP fallbacks and corrected FP16 variant gating, plus an internal field rename.

Improvements

  • Added fallback definitions for aarch64 HWCAP bits so builds against older kernel headers still compile and detect CPU features.
  • FP16 CPU variants on aarch64 now require HWCAP_ASIMDHP, preventing selection of half-precision kernels on CPUs that lack the vector extension.
  • Renamed the internal has_fp16_va field to has_fp16 to reflect that it gates the entire FEAT_FP16 extension, including scalar arithmetic.
  • Carries forward the prior Windows MSYS2 UCRT64 (GCC 16.1.0) crash mitigation from b10308.

Bug fixes

  • Corrected misdetection of the aarch64 FP16 variant that could enable unsupported half-precision code paths.

Upgrade notes

No configuration or model changes are required; this is a drop-in replacement for b10308. ARM64 users should re-check which CPU variant is selected at startup, since FP16 kernels may now be disabled on hardware lacking ASIMDHP. Only code that directly references ggml internals needs updating for the has_fp16_va to has_fp16 rename.

Analyzed Oct 30, 2026

Upgrading?

Release notes summary

This release improves ARM64 CPU support by adding fallback definitions for missing HWCAP bits and fixing the detection of the FP16 variant. It also renames has_fp16_va to has_fp16 for clarity.

Key Highlights

  • Added fallback definitions for missing aarch64 HWCAP bits (#25554)
  • Fixed FP16 variant detection by requiring HWCAP_ASIMDHP
  • Renamed has_fp16_va to has_fp16 to gate the whole FEAT_FP16 extension

Release Notes (excerpt)

<details open> ggml : add aarch64 HWCAP fallbacks and fix fp16 variant detection (#25554) * ggml : add fallback definitions for missing aarch64 HWCAP bits * ggml : require HWCAP_ASIMDHP for the aarch64 fp16 cpu variants Also rename has_fp16_va to has_fp16, the field gates the whole FEAT_FP16 extension, scalar and vector half-precision arithmetic together. </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories