Class SpeculativeDecoding

java.lang.Object

smile.llm.engine.SpeculativeDecoding

public final class SpeculativeDecoding
extends Object

Greedy speculative accept/reject helpers (Leviathan-style prefix match).

Given draft tokens d[0..N) and target samples t[0..N] (length N+1, where t[i] is the target sample at draft depth i and t[N] is the bonus after a full accept), accept the longest prefix where t[i] == d[i] and always emit the bonus token t[r] for r accepted drafts.

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static final record	SpeculativeDecoding.AcceptResult	Result of one speculative verify round.

Method Summary

Modifier and Type	Method	Description
static SpeculativeDecoding.AcceptResult	acceptGreedy(int[] drafts, int[] targetSamples)	Greedy accept: exact match between draft and target samples.

Methods inherited from class Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Details

acceptGreedy

public static SpeculativeDecoding.AcceptResult acceptGreedy(int[] drafts,
 int[] targetSamples)

Greedy accept: exact match between draft and target samples.

Parameters:

drafts - draft token ids [N] (N >= 1).

targetSamples - target samples [N+1] (greedy argmax per verify position).

Returns:

accept result with r+1 tokens.