The ggml‑org team released llama.cpp version b11528 on 2026‑10‑09. The update lets the library treat views of host‑allocated tensors as normal nodes, stopping a previous assert error that appeared with the –split-mode tensor option. It also brings back SM tensor support for the K2 Horizon accelerator. The change is included in the official GitHub release and applies to all listed platforms, from macOS to Windows and Linux.
Why it matters
Developers can now run llama.cpp with split‑mode KV caching on more hardware without crashes, expanding usable configurations for local LLM inference.
