b10299

ggml-org/llama.cppb10299Aug 7, 2026by github-actions[bot]same day as b10298

What changed since b10298

low risk

b10299 is a single Metal backend build fix for the lightning indexer kernel, with no functional changes beyond b10298.

Improvements

  • Metal kernel_lightning_indexer now declares a POD threadgroup array and casts it to a matrix pointer, sidestepping an MSL constructor limitation.
  • Metal builds that previously failed to compile this kernel should now build cleanly.

Bug fixes

  • Resolved a Metal shader compilation error ('no matching constructor') caused by instantiating an array of matrix types in threadgroup memory.

Upgrade notes

This is a build-only fix, so no config or model changes are required. Anyone on b10298 with a Metal build failure in kernel_lightning_indexer should move to b10299; others can upgrade without risk. The mtmd chunk save/load API and constant renames introduced in b10298 remain unchanged.

Analyzed Oct 30, 2026

Upgrading?

Release notes summary

This release fixes a compilation error in the Metal backend where declaring arrays of matrix types in `threadgroup` memory was not supported. The fix uses a POD array and casting.

Key Highlights

  • Fixed compilation error in Metal kernel `kernel_lightning_indexer`.
  • Workaround for MSL limitation regarding matrix array initialization in `threadgroup` memory.
  • Uses POD array and casting instead of direct matrix array declaration.

Release Notes (excerpt)

<details open> metal : avoid `threadgroup` matrix array instantiation in kernel_lightning_indexer (#26646) - In MSL, declaring an array of matrix types like `threadgroup half4x4` causes a 'no matching constructor' compilation error because MSL matrix types do not have zero-argument default constructors and threadgroup variables cannot have initializers. - Fix this by declaring a POD `threadgroup half` array instead and casting to `threadgroup half4x4 *` for matrix indexing.…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories