b9878

ggml-org/llama.cppb9878Jul 5, 2026by github-actions[bot]same day as b9877

What changed since b9877

low risk

Patch release fixes stale tensor-split metadata for draft models with GQA attention and reverts two recent internal commits.

Improvements

  • Tensor-split metadata for GQA attention in draft models is now computed correctly instead of carrying stale values.
  • Reverted the ggml-backend-meta asset hack, restoring the prior backend metadata handling.
  • Reverted the earlier code-tidying commit to keep style consistent with existing code.

Bug fixes

  • Draft models no longer inherit stale tensor-split parameters, which could misconfigure speculative decoding splits.
  • Corrected GQA attention tensor-split metadata that was previously written with wrong values.

Upgrade notes

This is a small follow-up to b9877 with no config or API changes; users running draft/speculative decoding with GQA models and custom tensor splits should upgrade to get correct split metadata. The multi-buffer abort introduced in b9877 remains in effect, so verify your backend buffer setup still starts cleanly.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

This release fixes stale tensor-split parameters for draft models, specifically correcting metadata for GQA attention. It also includes a revert of a recent code tidying commit and a backend meta asset hack.

Key Highlights

  • Fix stale tensor-split params for draft models
  • Fix GQA attention tensor split metadata
  • Revert code tidying changes
  • Revert ggml-backend-meta asset hack

Release Notes (excerpt)

<details open> Fix stale tensor-split params for draft models (#24814) * meta: fix tensor split metadata for GQA attention * Tidied the code a bit to match existing style * Revert "Tidied the code a bit to match existing style" This reverts commit b90c6c6300091fe09e2350a3d4edcfcf15db8d2e. * Reverted the ggml-backend-meta asset hack. </details> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories