llama.cpp adds new OpenCL kernels for small‑batch workloads

The ggml‑org/llama.cpp project released an update (b11763) on 2026‑10‑21. It adds a handful of OpenCL kernels that speed up small‑batch matrix multiplications, including q4_k dp4a GEMM and a narrow mul_mm_f16_f32 variant. The changes also fix an out‑of‑bounds read bug and remove an unnecessary CPU‑fallback guard. The library already runs on many platforms, from macOS Apple Silicon to Windows CUDA and Linux Vulkan, and now better supports Snapdragon CPUs, Adreno GPUs and Hexagon NPUs.