TL;DR: llama.cpp's OpenCL backend now uses cooperative-k dp4a GEMM kernels for Q4_0, Q4_K and Q6_K, speeding up quantized inference on OpenCL GPUs.
Summary: Build b11741 of ggerganov/llama.cpp (PR #30741) introduces cooperative-k dp4a GEMM implementations for the Q4_0, Q4_K and Q6_K quantization formats in the OpenCL backend, plus an alignment fix. The cooperative-k approach combines partial dot products across work-items using integer dp4a instructions, which are the fast path for 4-bit and 6-bit quantized matmuls on GPUs. This extends the OpenCL path's coverage of the most commonly used K-quant and legacy 4-bit formats.
Why it matters: OpenCL is the only GPU acceleration path on many AMD, Intel and mobile/embedded GPUs, so faster dp4a kernels directly raise local LLM throughput on hardware that CUDA and Metal don't reach. Worth benchmarking your Q4_0/Q4_K/Q6_K models against the previous build to measure the gain.
Source: github_releases