b9969

ggml-org/llama.cppb9969Jul 12, 2026by github-actions[bot]same day as b9968

What changed since b9968

low risk

Vulkan Adreno fix routes large matmuls to medium tiles, resolving llama-cli crashes on longer prompts.

Improvements

  • Large matmuls on Adreno GPUs now use the medium tile path, reducing shared memory pressure.
  • Small-pipeline matmul implementation was updated alongside the tiling change.
  • An unused Adreno device entry was removed from the Vulkan backend.

Bug fixes

  • llama-cli no longer crashes on longer prompt sizes with q4_0 quantized models on Adreno Vulkan, which was caused by insufficient shared memory.

Upgrade notes

This is a targeted Vulkan/Adreno stability patch with no config or API changes, so a straight binary swap is sufficient. Users running q4_0 models on Adreno GPUs via Vulkan should upgrade to avoid crashes on long prompts. The OpenCL int8/MoE Adreno work from b9968 is unaffected.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

Fixed a critical crash bug in llama-cli for longer prompts by updating Vulkan matmul routing for Adreno GPUs to use medium tiles and fixing shared memory issues.

Key Highlights

  • Fixed llama-cli crashes on longer prompts due to insufficient shared memory
  • Routed large matmuls to medium tile on Adreno GPUs
  • Removed unused Adreno device
  • Updated matmul implementation for small pipeline

New Features

  • Vulkan Adreno stability fix

Release Notes (excerpt)

<details open> Vulkan: route large matmuls to medium tile on Adreno (#24877) * [Vulkan] Fixes llama-cli breaking over longer promts sizes The llama-cli was breaking for longer promts sizes for q4_0 quantized networks. Causing due to insufficient shared memory. * Removed the un-used Adreno device * Updated matmul for small pipeline. </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