luboslenco
575 lines
14 KiB
C
575 lines
14 KiB
C
/*
|
|
* Iris Tokenizer Implementation
|
|
*
|
|
* BPE (Byte Pair Encoding) tokenizer for text-to-image generation.
|
|
* Supports both WordPiece-style and SentencePiece-style tokenization.
|
|
*/
|
|
|
|
#include "iris.h"
|
|
#include <ctype.h>
|
|
#include <stdio.h>
|
|
#include <stdlib.h>
|
|
#include <string.h>
|
|
|
|
/* ========================================================================
|
|
* Internal Data Structures
|
|
* ======================================================================== */
|
|
|
|
/* Hash table entry for vocabulary lookup */
|
|
typedef struct {
|
|
char *token;
|
|
int id;
|
|
} vocab_entry_t;
|
|
|
|
/* BPE merge rule */
|
|
typedef struct {
|
|
int left;
|
|
int right;
|
|
int result;
|
|
int priority; /* Lower = higher priority (earlier merge) */
|
|
} bpe_merge_t;
|
|
|
|
/* Tokenizer context */
|
|
struct iris_tokenizer {
|
|
/* Vocabulary */
|
|
char **vocab; /* id -> token string */
|
|
int vocab_size;
|
|
vocab_entry_t *vocab_hash; /* Hash table for token -> id lookup */
|
|
int hash_size;
|
|
|
|
/* BPE merges */
|
|
bpe_merge_t *merges;
|
|
int num_merges;
|
|
|
|
/* Special token IDs */
|
|
int pad_id;
|
|
int unk_id;
|
|
int bos_id;
|
|
int eos_id;
|
|
|
|
/* Configuration */
|
|
int max_length;
|
|
int add_bos;
|
|
int add_eos;
|
|
};
|
|
|
|
/* Forward declarations */
|
|
void iris_tokenizer_free(iris_tokenizer *tok);
|
|
|
|
/* ========================================================================
|
|
* Hash Table Functions
|
|
* ======================================================================== */
|
|
|
|
/* FNV-1a hash */
|
|
static unsigned int hash_string(const char *str) {
|
|
unsigned int hash = 2166136261u;
|
|
while (*str) {
|
|
hash ^= (unsigned char)*str++;
|
|
hash *= 16777619u;
|
|
}
|
|
return hash;
|
|
}
|
|
|
|
/* Insert into hash table */
|
|
static void vocab_hash_insert(vocab_entry_t *table, int hash_size, const char *token, int id) {
|
|
unsigned int h = hash_string(token) % hash_size;
|
|
while (table[h].token != NULL) {
|
|
if (strcmp(table[h].token, token) == 0) {
|
|
return; /* Already exists */
|
|
}
|
|
h = (h + 1) % hash_size;
|
|
}
|
|
table[h].token = strdup(token);
|
|
table[h].id = id;
|
|
}
|
|
|
|
/* Lookup in hash table, returns -1 if not found */
|
|
static int vocab_hash_lookup(const vocab_entry_t *table, int hash_size, const char *token) {
|
|
unsigned int h = hash_string(token) % hash_size;
|
|
while (table[h].token != NULL) {
|
|
if (strcmp(table[h].token, token) == 0) {
|
|
return table[h].id;
|
|
}
|
|
h = (h + 1) % hash_size;
|
|
}
|
|
return -1;
|
|
}
|
|
|
|
/* ========================================================================
|
|
* Unicode Utilities
|
|
* ======================================================================== */
|
|
|
|
/* Get UTF-8 character length from first byte */
|
|
static int utf8_char_len(unsigned char c) {
|
|
if ((c & 0x80) == 0)
|
|
return 1;
|
|
if ((c & 0xE0) == 0xC0)
|
|
return 2;
|
|
if ((c & 0xF0) == 0xE0)
|
|
return 3;
|
|
if ((c & 0xF8) == 0xF0)
|
|
return 4;
|
|
return 1; /* Invalid, treat as 1 */
|
|
}
|
|
|
|
/* Decode UTF-8 character to codepoint - reserved for QWEN3 text encoder */
|
|
__attribute__((unused)) static int utf8_decode(const char *s, int *len) {
|
|
unsigned char c = (unsigned char)s[0];
|
|
if ((c & 0x80) == 0) {
|
|
*len = 1;
|
|
return c;
|
|
}
|
|
if ((c & 0xE0) == 0xC0) {
|
|
*len = 2;
|
|
return ((c & 0x1F) << 6) | (s[1] & 0x3F);
|
|
}
|
|
if ((c & 0xF0) == 0xE0) {
|
|
*len = 3;
|
|
return ((c & 0x0F) << 12) | ((s[1] & 0x3F) << 6) | (s[2] & 0x3F);
|
|
}
|
|
if ((c & 0xF8) == 0xF0) {
|
|
*len = 4;
|
|
return ((c & 0x07) << 18) | ((s[1] & 0x3F) << 12) | ((s[2] & 0x3F) << 6) | (s[3] & 0x3F);
|
|
}
|
|
*len = 1;
|
|
return c;
|
|
}
|
|
|
|
/* Check if character is whitespace */
|
|
static int is_whitespace(int cp) {
|
|
return cp == ' ' || cp == '\t' || cp == '\n' || cp == '\r';
|
|
}
|
|
|
|
/* Check if character is punctuation */
|
|
static int is_punctuation(int cp) {
|
|
if (cp >= 33 && cp <= 47)
|
|
return 1; /* !"#$%&'()*+,-./ */
|
|
if (cp >= 58 && cp <= 64)
|
|
return 1; /* :;<=>?@ */
|
|
if (cp >= 91 && cp <= 96)
|
|
return 1; /* [\]^_` */
|
|
if (cp >= 123 && cp <= 126)
|
|
return 1; /* {|}~ */
|
|
return 0;
|
|
}
|
|
|
|
/* ========================================================================
|
|
* Pre-tokenization
|
|
* ======================================================================== */
|
|
|
|
/* Split text into initial tokens (words, whitespace, punctuation) */
|
|
static char **pretokenize(const char *text, int *num_tokens) {
|
|
int capacity = 64;
|
|
char **tokens = malloc(capacity * sizeof(char *));
|
|
int count = 0;
|
|
|
|
const char *p = text;
|
|
while (*p) {
|
|
/* Skip leading whitespace (but could keep as token) */
|
|
while (*p && is_whitespace((unsigned char)*p))
|
|
p++;
|
|
if (!*p)
|
|
break;
|
|
|
|
const char *start = p;
|
|
int char_len = utf8_char_len((unsigned char)*p);
|
|
|
|
if (is_punctuation((unsigned char)*p)) {
|
|
/* Single punctuation is a token */
|
|
p += char_len;
|
|
}
|
|
else {
|
|
/* Read until whitespace or punctuation */
|
|
while (*p && !is_whitespace((unsigned char)*p) && !is_punctuation((unsigned char)*p)) {
|
|
p += utf8_char_len((unsigned char)*p);
|
|
}
|
|
}
|
|
|
|
/* Create token */
|
|
int len = p - start;
|
|
char *token = malloc(len + 1);
|
|
memcpy(token, start, len);
|
|
token[len] = '\0';
|
|
|
|
if (count >= capacity) {
|
|
capacity *= 2;
|
|
tokens = realloc(tokens, capacity * sizeof(char *));
|
|
}
|
|
tokens[count++] = token;
|
|
}
|
|
|
|
*num_tokens = count;
|
|
return tokens;
|
|
}
|
|
|
|
/* ========================================================================
|
|
* BPE Tokenization
|
|
* ======================================================================== */
|
|
|
|
/* Tokenize a single word using BPE */
|
|
static int *bpe_tokenize_word(iris_tokenizer *tok, const char *word, int *num_tokens) {
|
|
int len = strlen(word);
|
|
if (len == 0) {
|
|
*num_tokens = 0;
|
|
return NULL;
|
|
}
|
|
|
|
/* Start with character-level tokens */
|
|
int max_chars = len; /* UTF-8 can have multi-byte chars */
|
|
int *tokens = malloc(max_chars * sizeof(int));
|
|
int n = 0;
|
|
|
|
const char *p = word;
|
|
while (*p) {
|
|
int char_len = utf8_char_len((unsigned char)*p);
|
|
char buf[8];
|
|
memcpy(buf, p, char_len);
|
|
buf[char_len] = '\0';
|
|
|
|
int id = vocab_hash_lookup(tok->vocab_hash, tok->hash_size, buf);
|
|
if (id < 0) {
|
|
/* Unknown character, try byte fallback */
|
|
for (int i = 0; i < char_len; i++) {
|
|
char byte_buf[8];
|
|
snprintf(byte_buf, sizeof(byte_buf), "<0x%02X>", (unsigned char)p[i]);
|
|
id = vocab_hash_lookup(tok->vocab_hash, tok->hash_size, byte_buf);
|
|
if (id < 0)
|
|
id = tok->unk_id;
|
|
tokens[n++] = id;
|
|
}
|
|
}
|
|
else {
|
|
tokens[n++] = id;
|
|
}
|
|
p += char_len;
|
|
}
|
|
|
|
/* Apply BPE merges */
|
|
int changed = 1;
|
|
while (changed && n > 1) {
|
|
changed = 0;
|
|
int best_idx = -1;
|
|
int best_priority = tok->num_merges + 1;
|
|
|
|
/* Find best merge */
|
|
for (int i = 0; i < n - 1; i++) {
|
|
for (int m = 0; m < tok->num_merges; m++) {
|
|
if (tok->merges[m].left == tokens[i] && tok->merges[m].right == tokens[i + 1]) {
|
|
if (tok->merges[m].priority < best_priority) {
|
|
best_priority = tok->merges[m].priority;
|
|
best_idx = i;
|
|
}
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
|
|
/* Apply best merge */
|
|
if (best_idx >= 0) {
|
|
for (int m = 0; m < tok->num_merges; m++) {
|
|
if (tok->merges[m].left == tokens[best_idx] && tok->merges[m].right == tokens[best_idx + 1]) {
|
|
tokens[best_idx] = tok->merges[m].result;
|
|
/* Shift remaining tokens */
|
|
for (int i = best_idx + 1; i < n - 1; i++) {
|
|
tokens[i] = tokens[i + 1];
|
|
}
|
|
n--;
|
|
changed = 1;
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
*num_tokens = n;
|
|
return tokens;
|
|
}
|
|
|
|
/* ========================================================================
|
|
* Tokenizer Loading
|
|
* ======================================================================== */
|
|
|
|
#define TOK_MAGIC "FTOK"
|
|
|
|
iris_tokenizer *iris_tokenizer_load(const char *path) {
|
|
FILE *f = fopen(path, "rb");
|
|
if (!f) {
|
|
fprintf(stderr, "iris_tokenizer_load: cannot open %s\n", path);
|
|
return NULL;
|
|
}
|
|
|
|
/* Check magic */
|
|
char magic[4];
|
|
if (fread(magic, 1, 4, f) != 4 || memcmp(magic, TOK_MAGIC, 4) != 0) {
|
|
fprintf(stderr, "iris_tokenizer_load: invalid magic\n");
|
|
fclose(f);
|
|
return NULL;
|
|
}
|
|
|
|
iris_tokenizer *tok = calloc(1, sizeof(iris_tokenizer));
|
|
if (!tok) {
|
|
fclose(f);
|
|
return NULL;
|
|
}
|
|
|
|
/* Read config */
|
|
uint32_t config[8];
|
|
if (fread(config, sizeof(uint32_t), 8, f) != 8)
|
|
goto error;
|
|
|
|
tok->vocab_size = config[0];
|
|
tok->num_merges = config[1];
|
|
tok->pad_id = config[2];
|
|
tok->unk_id = config[3];
|
|
tok->bos_id = config[4];
|
|
tok->eos_id = config[5];
|
|
tok->max_length = config[6];
|
|
tok->add_bos = config[7] & 1;
|
|
tok->add_eos = (config[7] >> 1) & 1;
|
|
|
|
tok->hash_size = tok->vocab_size * 2 + 1;
|
|
if (tok->hash_size < IRIS_VOCAB_HASH_SIZE)
|
|
tok->hash_size = IRIS_VOCAB_HASH_SIZE;
|
|
|
|
/* Allocate vocabulary */
|
|
tok->vocab = malloc(tok->vocab_size * sizeof(char *));
|
|
tok->vocab_hash = calloc(tok->hash_size, sizeof(vocab_entry_t));
|
|
if (!tok->vocab || !tok->vocab_hash)
|
|
goto error;
|
|
|
|
/* Read vocabulary */
|
|
for (int i = 0; i < tok->vocab_size; i++) {
|
|
uint16_t len;
|
|
if (fread(&len, sizeof(uint16_t), 1, f) != 1)
|
|
goto error;
|
|
|
|
tok->vocab[i] = malloc(len + 1);
|
|
if (!tok->vocab[i])
|
|
goto error;
|
|
if (fread(tok->vocab[i], 1, len, f) != len)
|
|
goto error;
|
|
tok->vocab[i][len] = '\0';
|
|
|
|
vocab_hash_insert(tok->vocab_hash, tok->hash_size, tok->vocab[i], i);
|
|
}
|
|
|
|
/* Read merges */
|
|
if (tok->num_merges > 0) {
|
|
tok->merges = malloc(tok->num_merges * sizeof(bpe_merge_t));
|
|
if (!tok->merges)
|
|
goto error;
|
|
|
|
for (int i = 0; i < tok->num_merges; i++) {
|
|
uint32_t merge[3];
|
|
if (fread(merge, sizeof(uint32_t), 3, f) != 3)
|
|
goto error;
|
|
tok->merges[i].left = merge[0];
|
|
tok->merges[i].right = merge[1];
|
|
tok->merges[i].result = merge[2];
|
|
tok->merges[i].priority = i;
|
|
}
|
|
}
|
|
|
|
fclose(f);
|
|
return tok;
|
|
|
|
error:
|
|
fprintf(stderr, "iris_tokenizer_load: error reading tokenizer\n");
|
|
fclose(f);
|
|
iris_tokenizer_free(tok);
|
|
return NULL;
|
|
}
|
|
|
|
void iris_tokenizer_free(iris_tokenizer *tok) {
|
|
if (!tok)
|
|
return;
|
|
|
|
if (tok->vocab) {
|
|
for (int i = 0; i < tok->vocab_size; i++) {
|
|
free(tok->vocab[i]);
|
|
}
|
|
free(tok->vocab);
|
|
}
|
|
|
|
if (tok->vocab_hash) {
|
|
for (int i = 0; i < tok->hash_size; i++) {
|
|
free(tok->vocab_hash[i].token);
|
|
}
|
|
free(tok->vocab_hash);
|
|
}
|
|
|
|
free(tok->merges);
|
|
free(tok);
|
|
}
|
|
|
|
/* ========================================================================
|
|
* Main Tokenization API
|
|
* ======================================================================== */
|
|
|
|
int *iris_tokenize(iris_tokenizer *tok, const char *text, int *num_tokens, int max_len) {
|
|
if (!tok || !text) {
|
|
*num_tokens = 0;
|
|
return NULL;
|
|
}
|
|
|
|
if (max_len <= 0)
|
|
max_len = tok->max_length;
|
|
|
|
/* Pre-tokenize */
|
|
int num_words;
|
|
char **words = pretokenize(text, &num_words);
|
|
|
|
/* Collect all tokens */
|
|
int capacity = 128;
|
|
int *all_tokens = malloc(capacity * sizeof(int));
|
|
int total = 0;
|
|
|
|
/* Add BOS if configured */
|
|
if (tok->add_bos && tok->bos_id >= 0) {
|
|
all_tokens[total++] = tok->bos_id;
|
|
}
|
|
|
|
/* BPE tokenize each word */
|
|
for (int w = 0; w < num_words && total < max_len - (tok->add_eos ? 1 : 0); w++) {
|
|
int n;
|
|
int *word_tokens = bpe_tokenize_word(tok, words[w], &n);
|
|
|
|
for (int i = 0; i < n && total < max_len - (tok->add_eos ? 1 : 0); i++) {
|
|
if (total >= capacity) {
|
|
capacity *= 2;
|
|
all_tokens = realloc(all_tokens, capacity * sizeof(int));
|
|
}
|
|
all_tokens[total++] = word_tokens[i];
|
|
}
|
|
|
|
free(word_tokens);
|
|
free(words[w]);
|
|
}
|
|
free(words);
|
|
|
|
/* Add EOS if configured */
|
|
if (tok->add_eos && tok->eos_id >= 0 && total < max_len) {
|
|
if (total >= capacity) {
|
|
capacity *= 2;
|
|
all_tokens = realloc(all_tokens, capacity * sizeof(int));
|
|
}
|
|
all_tokens[total++] = tok->eos_id;
|
|
}
|
|
|
|
*num_tokens = total;
|
|
return all_tokens;
|
|
}
|
|
|
|
/* Decode tokens back to text */
|
|
char *iris_detokenize(iris_tokenizer *tok, const int *tokens, int num_tokens) {
|
|
if (!tok || !tokens || num_tokens <= 0) {
|
|
return strdup("");
|
|
}
|
|
|
|
/* Calculate total length */
|
|
int total_len = 0;
|
|
for (int i = 0; i < num_tokens; i++) {
|
|
int id = tokens[i];
|
|
if (id >= 0 && id < tok->vocab_size) {
|
|
total_len += strlen(tok->vocab[id]);
|
|
}
|
|
}
|
|
|
|
/* Build string */
|
|
char *text = malloc(total_len + 1);
|
|
char *p = text;
|
|
for (int i = 0; i < num_tokens; i++) {
|
|
int id = tokens[i];
|
|
if (id >= 0 && id < tok->vocab_size) {
|
|
/* Skip special tokens */
|
|
if (id == tok->bos_id || id == tok->eos_id || id == tok->pad_id) {
|
|
continue;
|
|
}
|
|
const char *token = tok->vocab[id];
|
|
int len = strlen(token);
|
|
memcpy(p, token, len);
|
|
p += len;
|
|
}
|
|
}
|
|
*p = '\0';
|
|
|
|
return text;
|
|
}
|
|
|
|
/* Get vocabulary size */
|
|
int iris_tokenizer_vocab_size(iris_tokenizer *tok) {
|
|
return tok ? tok->vocab_size : 0;
|
|
}
|
|
|
|
/* Get token string by ID */
|
|
const char *iris_tokenizer_get_token(iris_tokenizer *tok, int id) {
|
|
if (!tok || id < 0 || id >= tok->vocab_size)
|
|
return NULL;
|
|
return tok->vocab[id];
|
|
}
|
|
|
|
/* Get token ID by string */
|
|
int iris_tokenizer_get_id(iris_tokenizer *tok, const char *token) {
|
|
if (!tok || !token)
|
|
return -1;
|
|
return vocab_hash_lookup(tok->vocab_hash, tok->hash_size, token);
|
|
}
|
|
|
|
/* ========================================================================
|
|
* Simple Tokenizer (Fallback)
|
|
* ======================================================================== */
|
|
|
|
/*
|
|
* Create a simple tokenizer that just does character-level tokenization.
|
|
* Used as fallback or for testing.
|
|
*/
|
|
iris_tokenizer *iris_tokenizer_create_simple(void) {
|
|
iris_tokenizer *tok = calloc(1, sizeof(iris_tokenizer));
|
|
if (!tok)
|
|
return NULL;
|
|
|
|
/* ASCII printable characters + some special tokens */
|
|
tok->vocab_size = 256 + 4; /* All bytes + special tokens */
|
|
tok->hash_size = 521; /* Prime > vocab_size */
|
|
tok->num_merges = 0;
|
|
tok->pad_id = 256;
|
|
tok->unk_id = 257;
|
|
tok->bos_id = 258;
|
|
tok->eos_id = 259;
|
|
tok->max_length = IRIS_MAX_SEQ_LEN;
|
|
tok->add_bos = 1;
|
|
tok->add_eos = 1;
|
|
|
|
tok->vocab = malloc(tok->vocab_size * sizeof(char *));
|
|
tok->vocab_hash = calloc(tok->hash_size, sizeof(vocab_entry_t));
|
|
if (!tok->vocab || !tok->vocab_hash) {
|
|
iris_tokenizer_free(tok);
|
|
return NULL;
|
|
}
|
|
|
|
/* Create vocabulary */
|
|
for (int i = 0; i < 256; i++) {
|
|
char buf[8];
|
|
if (i >= 32 && i < 127) {
|
|
buf[0] = (char)i;
|
|
buf[1] = '\0';
|
|
}
|
|
else {
|
|
snprintf(buf, sizeof(buf), "<0x%02X>", i);
|
|
}
|
|
tok->vocab[i] = strdup(buf);
|
|
vocab_hash_insert(tok->vocab_hash, tok->hash_size, tok->vocab[i], i);
|
|
}
|
|
tok->vocab[256] = strdup("<pad>");
|
|
tok->vocab[257] = strdup("<unk>");
|
|
tok->vocab[258] = strdup("<bos>");
|
|
tok->vocab[259] = strdup("<eos>");
|
|
|
|
vocab_hash_insert(tok->vocab_hash, tok->hash_size, "<pad>", 256);
|
|
vocab_hash_insert(tok->vocab_hash, tok->hash_size, "<unk>", 257);
|
|
vocab_hash_insert(tok->vocab_hash, tok->hash_size, "<bos>", 258);
|
|
vocab_hash_insert(tok->vocab_hash, tok->hash_size, "<eos>", 259);
|
|
|
|
return tok;
|
|
}
|