Tools
llama.cpp b10941 fixes an OpenCL quantization alignment bug that had spread from q6_K to three other formats
Build b10941, tagged 2026-09-13T10:34Z, merges #28575, which applies the noshuffle row-alignment rule already used for q6_K to q4_K, q5_K and q8_0 on the OpenCL backend, a correctness fix affecting anyone running those quants on OpenCL hardware. The same window merges #28830, a workaround for an NVIDIA queuesubmit driver bug in the Vulkan backend, and #28742, which improves parsing of complex types in qwen3-coder chat templates. Only six PRs merged in the 30 hours to that tag, three of them CI changes.
Source
↳ Follow the thread