b10058
ggml-org/llama.cppb10058Jul 17, 2026by github-actions[bot]same day as b10057
What changed since b10057
b10058 adds Vulkan Q2_0 quantization support and speeds up mat-vec-mul, on top of b10057's SYCL fixes.
New features
- Vulkan backend can now run models quantized with the Q2_0 format.
- Workgroup reordering introduced in the Vulkan mat-vec-mul path to improve memory access patterns.
Improvements
- Vulkan mat-vec-mul throughput improved by doubling the number of rows processed per workgroup.
- Error threshold for the f16-to-Q2_0 set_rows test adjusted to match the new kernel behavior.
Bug fixes
- Carried over from b10057: SYCL row calculation corrected when K_QUANTS_PER_ITERATION is 1.
- Carried over from b10057: SYCL Q5_K kernels now consistently use K_QUANTS_PER_ITERATION, including the non-reordered variant.
- Carried over from b10057: missing second-half processing in the reordered Q5_K kernel and a potential off-by-one / row-bound check fixed.
Upgrade notes
This is an additive backend change: existing CPU, CUDA, SYCL and Vulkan workflows are unaffected, so no config or model migration is needed. Users on Vulkan who want to try Q2_0 quantized models should pull the new Vulkan build artifacts. SYCL users should still take b10057 or later to get the quantization kernel fixes.
Analyzed Oct 31, 2026
Upgrading?
Release notes summary
Adds support for Q2_0 quantization format to the Vulkan backend and optimizes mat-vec-mul performance.
Key Highlights
- Vulkan backend now supports Q2_0 quantization
- Optimized mat-vec-mul performance by doubling rows per workgroup
- Implemented workgroup reordering
Release Notes (excerpt)
<details open> vulkan: Support Q2_0 (#25430) * vulkan: Support Q2_0 The backend perf tests for mat-vec-mul weren't very good at first (worse than q2_k), doubling the rows per workgroup made a big difference. * reorder * resolve merge conflict, adjust err threshold for f16->q2_0 set_rows </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:**…
Published by ggml-org on GitHub.
Read the full release notesRelated Repositories
:electron: Build cross-platform desktop apps with JavaScript, HTML, and CSS
Godot Engine – Multi-platform 2D and 3D game engine
ClickHouse® is a real-time analytics database management system
Self-hosted game stream host for Moonlight.
DuckDB is a high-performance analytical database system