Livt.ML

src/transformer/IIncrementalTransformerKernels.lvt

namespace Livt.ML.Transformer

/**
 * Extends existing kernels without changing the encoder interface.
 * CachedAttention consumes one query and token-major K/V with distinct length.
 * A negative-one bias tensor disables learned bias (cross attention).
 * Buckets are word entries for relative distances -(MAX_TOKENS-1)..+(MAX_TOKENS-1).
 * ProjectArgMax streams rounded/clipped logits; lowest row wins ties; -1 is error.
 */
interface IIncrementalTransformerKernels : ITransformerKernels
{
	/**
	 * Computes one query against length token-major cached K/V vectors. Width is features per
	 * head; position is the query token position. Causal mode excludes keys beyond position;
	 * zero masks exclude keys. Tensor -1 disables bias, otherwise buckets holds 2*MAX_TOKENS-1
	 * entries centered on relative position zero. Target must not overlap any consumed input
	 * span. True means completed writes; false may leave partial output.
	 */
	fn CachedAttention(tensor: int, q: int, k: int, v: int, target: int, masks: int,
		length: int, heads: int, width: int, position: int, causal: bool, buckets: int) bool

	/**
	 * Projects one columns-element vector through rows weight rows without storing logits.
	 * Returns the lowest row index among equal rounded/clipped maxima, or -1 on invalid shape,
	 * storage, or weights. Does not write activation memory.
	 */
	fn ProjectArgMax(tensor: int, source: int, rows: int, columns: int) int
}