b10189

ggml-org/llama.cppb10189Jul 30, 2026by github-actions[bot]same day as b10188

What changed since b10188

low risk

b10189 refactors the Apple M3 graph to use stock ops instead of a custom CPU op, on top of b10188's Metal memory-leak fix.

New features

  • M3 (Apple Silicon) graph now expresses its previously custom CPU operation using standard internal ops.

Improvements

  • Removal of the bespoke CPU op in the M3 graph simplifies the backend and reduces custom code to maintain.
  • Metal memory unwiring is now handled correctly when a model is freed before any GPU work runs (carried from b10188).
  • Metal regression test is always built and now measures system-wide wired memory.

Bug fixes

  • Fixed a Metal backend memory leak occurring when a model was freed without any GPU operations.

Upgrade notes

This is a routine build-to-build bump with no config or API changes; binaries can be swapped directly. Apple Silicon/Metal users get both the leak fix and the M3 graph refactor, so re-run any Metal-specific benchmarks to confirm performance is unchanged.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Refactors the M3 (Apple Silicon) graph by removing a custom CPU operation and implementing it using standard internal operations.

Key Highlights

  • M3 graph optimization
  • Replacement of custom ops with stock ops

Release Notes (excerpt)

<details open> Remove custom cpu op from the M3 graph, express with stock ops (#26297) </details> **Website:** - <https://llama.app> **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)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories