The ggml-org team released llama.cpp b12171 with a Vulkan change: it detects whether a system supports ReBAR, or Resizable BAR, and avoids BAR allocations when support is unavailable. The release page lists downloadable builds for macOS, iOS and Linux, including CPU and Vulkan versions for some platforms.
The release notes do not say which hardware or workloads will see a practical difference.
Why it matters
For people running llama.cpp with Vulkan on systems without ReBAR support, the change is intended to avoid BAR allocations; the release notes do not quantify its effect on performance or memory use.
