Livt.ML

src/transformer/ITransformerKernels.lvt

namespace Livt.ML.Transformer

/**
 * Serialized tensor operations. All storage addresses count int32 words, not bytes.
 * False invalidates the request: destinations may contain partial results.
 * Providers and source spans must remain stable until an operation completes.
 */
interface ITransformerKernels
{
	/**
	 * Copies length embedding rows selected by token IDs into a length-by-width target.
	 * Addresses count int32 words; target must not overlap the token span. Returns true after
	 * all writes; false may leave partial output.
	 */
	fn Embed(tensor: int, tokens: int, target: int, length: int, width: int) bool

	/**
	 * Applies affine RMS normalization to length vectors of width elements, using a 1-by-width
	 * weight tensor. Source and target word spans must be disjoint. True means completed
	 * writes; false may leave partial output.
	 */
	fn Norm(tensor: int, source: int, target: int, length: int, width: int) bool

	/**
	 * Projects length token-major vectors through a rows-by-columns weight tensor. Source and
	 * target word spans must be disjoint. Returns true after all rows are written; false may
	 * leave partial output.
	 */
	fn Linear(tensor: int, source: int, target: int, length: int, rows: int, columns: int) bool

	/**
	 * Computes unscaled attention over token-major Q/K/V, with heads groups of width features
	 * and nonzero key masks. Uses learned relative-position bias. Target must be disjoint from
	 * Q/K/V and masks; at least one key must be enabled. True means completed writes; false
	 * may leave partial output.
	 */
	fn Attention(tensor: int, q: int, k: int, v: int, target: int, masks: int,
		length: int, heads: int, width: int) bool

	/**
	 * Writes GELU(a) multiplied by b for count fixed-point elements. All addresses count int32
	 * words; target must be disjoint from both inputs. True means completed writes; false may
	 * leave partial output.
	 */
	fn Gate(a: int, b: int, target: int, count: int) bool

	/**
	 * Writes count saturated elementwise sums. Target may equal either input base, but partial
	 * overlap is rejected. Addresses count int32 words. True means completed writes; false may
	 * leave partial output.
	 */
	fn Add(a: int, b: int, target: int, count: int) bool
}