TL;DR: Version b10007 of llama.pp fixes an OpenCL bug that prevented backend initialization on devices without cl_khr_integer_dot_product, improving cross-hardware compatibility.
Summary: The b10007 release of llama.cpp fixes a DP4A bug where backend init would fail on OpenCL devices missing the cl_khr_integer_dot_product extension. The patch avoids calling dp4 kernels when the extension is unavailable, allowing the backend to initialize and fall back to alternative paths. This improves support for older or limited GPUs and mobile devices.
Why it matters: Indie developers running LLMs on diverse or edge hardware gain broader compatibility without crashing. If you deploy on OpenCL devices (e.g., older GPUs, some mobile SoCs), upgrade to b10007 to avoid init failures.
Source: github_releases