ggml-org released llama.cpp b11895 with a fix for overflow in large matrix output offsets on Metal. The change also widens batch offsets and token indices used by the MUL_MAT_ID operation.
The release page provides downloads for Apple Silicon and Intel Macs, iOS, and several Ubuntu CPU and Vulkan builds.
Why it matters
Developers running llama.cpp on Metal may avoid offset overflow in large matrix outputs, though the release notes do not say how often the issue occurred or which workloads trigger it.
