b10244

ggml-org/llama.cppb10244Aug 3, 2026by github-actions[bot]same day as b10243

What changed since b10243

low risk

b10244 moves M3 MSA attention logic into a new llama-kv-cache-msa module, with minor workspace fixes.

Breaking changes

  • Source-level reorganization splits MSA code out of llama-kv-cache into a new llama-kv-cache-msa file, which can break out-of-tree builds or patches that reference the old location.

New features

  • Dedicated memory implementation module for M3 MSA, separating attention memory handling from the general KV cache code.

Improvements

  • Cleaner separation of concerns between generic KV cache and M3-specific MSA memory handling.
  • Minor workspace (ws) and continuity corrections folded into the refactor.

Bug fixes

  • Small workspace-related issues addressed as part of the MSA move.

Upgrade notes

No user-facing config or CLI changes are indicated; a rebuild is sufficient. If you vendor or patch llama.cpp sources, update any references to the old MSA location in llama-kv-cache. Validate M3 model output after upgrading, since the attention memory path was restructured.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Refactored the MSA (Multi-Head Attention) logic for the M3 model into a dedicated memory implementation module.

Key Highlights

  • Moved MSA logic from llama-kv-cache to llama-kv-cache-msa
  • Minor fixes for workspace (ws) and continuity

Release Notes (excerpt)

<details open> model: M3: Move MSA into a new memory implementation (#26338) * Move MSA logic from llama-kv-cache into llama-kv-cache-msa * cont : minor * cont : ws fix --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> </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)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories