b9550

ggml-org/llama.cppb9550Jun 7, 2026by github-actions[bot]same day as b9549

What changed since b9549

low risk

b9550 is a targeted bug-fix release that corrects KV cache sharing to prevent tensor-overflow crashes during graph reservation.

Improvements

  • Shared KV cache cells now inherit the source cache size instead of assuming the draft default.
  • Graph reservation is more robust when the fitted target context is smaller than the draft default.

Bug fixes

  • Fixed a crash caused by oversized assistant views overflowing shared K/V tensors and tripping the ggml_view_4d size assert during graph reserve.

Upgrade notes

No configuration or API changes are required; this is a drop-in replacement for b9549. Users running speculative decoding or KV cache sharing with small fitted contexts should upgrade to avoid the reservation crash.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

This release fixes a crash/bug where oversized assistant views caused tensor overflow during graph reservation by ensuring the shared KV cache size follows the source cache size.

Key Highlights

  • Fixes crash due to tensor overflow during graph reserve.
  • Ensures shared KV cache size follows source cache size.
  • Resolves issues with fitted target contexts smaller than draft defaults.

New Features

  • Fixed KV cache sharing logic for small contexts

Release Notes (excerpt)

<details open> kv-cache: follow the source cache size when sharing cells (#24267) A fitted target context can end up smaller than the draft default, the oversized assistant views then overflow the shared K/V tensors and trip the ggml_view_4d size assert during graph reserve. </details> **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