For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo function

dispatch_relative_logits_mask

def dispatch_relative_logits_mask[dtype: DType, layout: Layout, //, callback_fn: def[mask_t: MHAMask](mask: mask_t) raises capturing thin -> None, local_window_size: Int = Int(-1)](bias_nd: LayoutTensor[dtype, layout], cache_lengths: LayoutTensor[DType.uint32, Layout.row_major(Int(-1)), ImmutAnyOrigin], input_row_offsets: LayoutTensor[DType.uint32, Layout.row_major(Int(-1)), ImmutAnyOrigin])

Wrap bias_nd in a RelativeLogitsMask and invoke callback_fn.

Like dispatch_materialized_mask, this carries runtime state (the bias table plus the tensors that recover its ragged-flat row), so it lives outside dispatch_mask's zero-arg string dispatch. local_window_size picks the visibility mask: <= 0 (canonically -1, the graph-level "no window" value) -> CausalMask, else SlidingWindowCausalMask[local_window_size]. The mask structs themselves report "no window" as sliding_window_size() == 0; this function is where the two conventions meet.