ggml-org’s llama.cpp b11971 release adds a tiled OpenCL q4_K matrix operation for Adreno GPUs. The change also routes the q4_K language-model output layer through that operation when the batch size is greater than one.
The release includes downloadable builds for macOS, iOS and several Linux platforms. The release notes do not give performance measurements, so the speed or resource impact of the Adreno change is unclear.
Why it matters
Developers running llama.cpp on Adreno hardware can try the new OpenCL path, but the release notes do not say whether it makes inference faster.
