Skip to main content

default_swa_pattern

Function default_swa_pattern 

Source
pub fn default_swa_pattern(arch: &str) -> Option<usize>
Expand description

The alternating sliding-window period an architecture uses when its GGUF carries {arch}.attention.sliding_window but not {arch}.attention.sliding_window_pattern.

The period is not in the file for these families — llama.cpp hardcodes it per architecture and only lets the metadata key override it (ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false) after seeding swa_period with the literal below). A missing key therefore does not mean “every layer is windowed”, which is what ferrox assumed: layer_sliding_window returns the window for all layers when swa_pattern is None, so a gpt-oss or cohere2 checkpoint ran its full-attention layers through a 128-token window and answered from a truncated history.

Values transcribed from each arch’s load_arch_hparams (src/models/*.cpp); None means “no per-arch default”, i.e. the window applies uniformly when one is declared.