CYBER WATCH / EQUINOXE INTELLIGENCE
← All articles
b12140: ggml: support `truncate=false` for YaRN (#30206)
llama.cpp Releases · 2026-11-06 02:01:33
ggml now supports the `truncate=false` option for YaRN, allowing the boundaries of correction intervals used in frequency interpolation to remain unrounded. This change is required for models that use `truncate=false`, such as gpt‑oss and some PLaMo‑3 models. GGML operations and the CPU backend have been updated to handle calculations without rounding, while the default remains `truncate=true`. Additional tests were added for the CUDA backend and RoPE helpers were moved to an internal header.Read original source ↗