# llama.cpp changes Hexagon flash-attention splitting

> Release b11868 combines two Hexagon flash-attention split flags and adds team-splitting when KV heads do not fully use available cores.

Oossa · 2026-10-24 · https://oossa.com/en/llama-cpp-changes-hexagon-flash-attention-splitting

The ggml-org/llama.cpp b11868 release changes how Hexagon handles multicore HMX flash attention. It replaces the separate `FA_HEAD_SPLIT` and `FA_SPLIT` settings with one flag, `GGML_HEXAGON_FA_SPLIT`.

With the flag set to 1, work can be split across KV heads and, when there are fewer KV heads than cores, across teams so otherwise idle cores can participate. The release gives Gemma 4 with two KV heads on four cores as an example.

## The facts

- `GGML_HEXAGON_FA_SPLIT` set to 0 keeps token-splitting only; 1 enables head-splitting and team-splitting when needed.
- The release notes say evenly distributed heads use plain head-splitting.

## Why it matters

For people running llama.cpp on Hexagon hardware, the change is intended to let cores participate when a model has fewer KV heads than cores; the release notes do not quantify any speedup.

## Sources & references

1. [ggml-org/llama.cpp b11868](https://github.com/ggml-org/llama.cpp/releases/tag/b11868) – llama.cpp, 2026-10-24

Last updated: 2026-10-24