b10197
ggml-org/llama.cppb10197Jul 30, 2026by github-actions[bot]same day as b10196
What changed since b10196
Adds alternative conv2d layout (cwhn) test support plus graph-build-time layout checks and a CPU kernel contiguity fix.
New features
- conv2d test cases can now exercise an alternative channel-width-height layout via a new cwhn flag.
- Graph construction now validates convolution layouts for both the generic and CUDA conv2d kernels.
- The Vulkan backend reports support for conv operations through a new build-time capability check.
Improvements
- CPU backend now stores convolution kernels contiguously, avoiding failures when the alternative layout is used.
- Vulkan device support queries reject unsupported conv layouts early, preventing run-time assertion failures.
- CI stability improved by catching layout mismatches at graph build time rather than at execution.
Bug fixes
- Fixed a Vulkan run-time assert failure triggered by unsupported convolution layouts.
- Fixed CPU backend test failures caused by non-contiguous kernel storage with cwhn=1.
Upgrade notes
This is a build/test-infrastructure oriented patch with no API or CLI changes; a straight rebuild is sufficient. Custom backends implementing conv2d should ensure kernels are stored contiguously and honor the layout check, since unsupported layouts are now rejected at graph build time.
Analyzed Oct 31, 2026
Upgrading?
Release notes summary
This release adds support for alternative convolution layouts in tests and fixes kernel storage requirements for the CPU backend to prevent test failures.
Key Highlights
- Added layout checks for `conv2d` kernel at graph build time
- Fixed CPU backend kernel storage for contiguous memory
- Added Vulkan backend support checks
Release Notes (excerpt)
<details open> Test support for alternative conv layout (#25617) * add bool cwhn = true to conv_2d test cases * add layout check at graph building time * extend layout checks for conv2d.cu kernel * in CPU back-end kernel needs to be stored contiguously to prevent test failures with cwhn=1 * trim white space * do op support check in vulkan backend * fix CI failure and vulkan run-time assert failure by introducing new graph build-time check in ggml_backend_vk_device_supports_op…
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