llama.cpp b10069: Adreno OpenCL broadcast support
llama.cpp build b10069 adds OpenCL broadcast support for Adreno MUL_MAT operations and fixes view offset handling for Adreno Q8_0 MUL_MAT in llama-server multi-stream mode.
b10069, released today, adds OpenCL broadcast support for Adreno MUL_MAT operations and fixes view_offs handling for Adreno Q8_0 MUL_MAT in llama-server multi-stream mode. The commit was co-authored by a Qualcomm contributor, which tracks given Adreno is Qualcomm’s GPU family. For anyone running llama.cpp on an Android device via the Windows arm64 OpenCL Adreno build, this is the build to grab. On-device inference on Snapdragon hardware has been a real use case for a while now, and steady OpenCL fixes are what make it reliable rather than experimental.