TL;DR: llama.cpp b10659 fixes Windows ROCm runtime loading by bundling HIP DLLs directly with the binaries, improving out-of-the-box AMD GPU support.
Summary: The b10659 release of llama.cpp includes a CI change that copies amdhip64_7, amd_comgr, and rocm_kpack next to the binaries, ensuring the correct HIP runtime loads over the driver's copy in System32. This fixes issue #26929, which caused runtime mismatches on Windows ROCm builds. The release also ships updated iOS XCFramework and UI bundles.
Why it matters: AI builders running local LLMs on AMD GPUs in Windows get a more reliable llama.cpp setup without manual DLL management. Update to b10659 if you target Windows ROCm, and test your existing AMD inference pipelines for improved stability.
Source: github_releases