TL;DR: llama.cpp significantly improves prefill performance for certain models on Adreno X2E GPUs by optimizing OpenCL matrix multiplication paths.
Summary: ggerganov/llama.cpp release b10687 introduces a performance improvement for Adreno X2E GPUs. It enables a faster xmem F16xF32 GEMM path for specific matrix multiplications, which was previously opt-in. This optimization bypasses a slower kernel, particularly benefiting models like gpt-oss-20b where non-expert weights remain f16.
Why it matters: Indie developers targeting mobile or edge AI with Adreno X2E GPUs can expect faster inference for certain models. Experiment with f16 models on Adreno X2E devices to leverage these performance gains.
Source: github_releases