b10188

ggml-org/llama.cppb10188Jul 30, 2026by github-actions[bot]same day as b10186

What changed since b10186

low risk

Metal backend now correctly unwires model memory when freed before any GPU work, fixing a leak.

Improvements

  • Metal backend releases wired memory properly on model teardown, even if no GPU operations ever ran.
  • Dummy GPU work is now only executed when residency sets are in use, avoiding unnecessary overhead.
  • A regression test for wired-memory behavior is now always built, improving coverage of the Metal path.

Bug fixes

  • Fixed a Metal memory leak that occurred when a model was freed without any prior GPU operations.

Upgrade notes

No API or configuration changes; this is a drop-in replacement for b10186. Metal/macOS users who repeatedly load and unload models without inference should see reduced wired-memory growth. Non-Metal backends are unaffected.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Fixes a memory leak in the Metal backend where memory was not properly released if the model was freed before any GPU operations occurred.

Key Highlights

  • Metal memory leak fix
  • Proper memory unwiring implementation

New Features

  • Metal backend memory management fix

Release Notes (excerpt)

<details open> metal: fix memory unwire if model is freed without any GPU operations (#26082) * metal: fix memory leak if model is freed without any GPU operations * metal: run dummy work only if residency sets are used * metal: wrap function in #if defined * metal: measure system-wide wired memory in test * metal: always build regression test Co-authored-by: YiChen Lv <63285796+forforever73@users.noreply.github.com> ---------…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories