The llama.cpp project released a new version (b11707) that adds a feature called GraniteSwitchDecisionFullOptions. It lets each switch model use up to 256 unique single‑token markers instead of the previous 66. The code automatically discovers which characters form single‑token entries for a given model’s vocabulary. This change is in the C++ conversion layer and is documented in the release notes.
Why it matters
Developers can build richer decision‑making prompts for Granite models without manually crafting token lists.
