1#pragma once
2
3#include "llama.h"
4
5#include <unordered_map>
6#include <string>
7#include <vector>
8
9#define LLAMA_NGRAM_MIN 1
10#define LLAMA_NGRAM_MAX 4
11#define LLAMA_NGRAM_STATIC 2
12
13// Data structures to map n-grams to empirical token probabilities:
14
15struct common_ngram {
16 llama_token tokens[LLAMA_NGRAM_MAX];
17
18 common_ngram() {
19 for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
20 tokens[i] = LLAMA_TOKEN_NULL;
21 }
22 }
23
24 common_ngram(const llama_token * input, const int ngram_size) {
25 for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
26 tokens[i] = i < ngram_size ? input[i] : LLAMA_TOKEN_NULL;
27 }
28 }
29
30 bool operator==(const common_ngram & other) const {
31 for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
32 if (tokens[i] != other.tokens[i]) {
33 return false;
34 }
35 }
36 return true;
37 }
38};
39
40struct common_token_hash_function {
41 size_t operator()(const llama_token token) const {
42 // see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
43 return token * 11400714819323198485llu;
44 }
45};
46
47struct common_ngram_hash_function {
48 size_t operator()(const common_ngram & ngram) const {
49 size_t hash = common_token_hash_function{}(ngram.tokens[0]);
50 for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) {
51 hash ^= common_token_hash_function{}(ngram.tokens[i]);
52 }
53 return hash;
54 }
55};
56
57// token -> number of times token has been seen
58typedef std::unordered_map<llama_token, int32_t> common_ngram_cache_part;
59
60// n-gram -> empirical distribution of following tokens
61typedef std::unordered_map<common_ngram, common_ngram_cache_part, common_ngram_hash_function> common_ngram_cache;
62
63
64// Update an ngram cache with tokens.
65// ngram_cache: the cache to modify.
66// ngram_min/ngram_max: the min/max size of the ngrams to extract from inp_data.
67// inp_data: the token sequence with which to update ngram_cache.
68// nnew: how many new tokens have been appended to inp_data since the last call to this function.
69// print_progress: whether to print progress to stderr.
70//
71// In order to get correct results inp_data can ONLY BE APPENDED TO.
72// Changes in the middle need a complete rebuild.
73void common_ngram_cache_update(
74 common_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector<llama_token> & inp_data, int nnew, bool print_progress);
75
76// Try to draft tokens from ngram caches.
77// inp: the tokens generated so far.
78// draft: the token sequence to draft. Expected to initially contain the previously sampled token.
79// n_draft: maximum number of tokens to add to draft.
80// ngram_min/gram_max: the min/max size of the ngrams in nc_context and nc_dynamic.
81// nc_context: ngram cache based on current context.
82// nc_dynamic: ngram cache based on previous user generations.
83// nc_static: ngram cache generated from a large text corpus, used for validation.
84void common_ngram_cache_draft(
85 std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
86 common_ngram_cache & nc_context, common_ngram_cache & nc_dynamic, common_ngram_cache & nc_static);
87
88// Save an ngram cache to a file.
89// ngram_cache: the ngram cache to save.
90// filename: the path under which to save the ngram cache.
91void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename);
92
93// Load an ngram cache saved with common_ngram_cache_save.
94// filename: the path from which to load the ngram cache.
95// returns: an ngram cache containing the information saved to filename.
96common_ngram_cache common_ngram_cache_load(std::string & filename);
97
98// Merge two ngram caches.
99// ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add.
100// ngram_cache_add: the ngram cache to add to ngram_cache_target.
101void common_ngram_cache_merge(common_ngram_cache & ngram_cache_target, common_ngram_cache & ngram_cache_add);
102