tokenizer.c File Reference

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>
#include <stdbool.h>
#include <ctype.h>
#include "tokenizer/tokenizer.h"
#include "tokenizer/murmurhash3.h"
#include "tokenizer/hash_table.h"

Go to the source code of this file.

Functions
int	ck_tokenizer_add_merge (CKTokenizer *tok, int32_t left, int32_t right, int32_t merged, int32_t priority)
int	ck_tokenizer_add_special_token (CKTokenizer *tok, const char *name, int32_t id)
int	ck_tokenizer_add_token (CKTokenizer *tok, const char *token, int32_t id, float score)
CKTokenizer *	ck_tokenizer_create (CKTokenizerType type)
int	ck_tokenizer_decode (const CKTokenizer *tok, const int32_t *ids, int num_ids, char *text, int max_len)
CKSpacePrefixStyle	ck_tokenizer_detect_space_prefix_style (CKTokenizer *tok)
int	ck_tokenizer_encode (const CKTokenizer *tok, const char *text, int text_len, int32_t *ids, int max_ids)
int	ck_tokenizer_encode_spm_dispatch (const CKTokenizer *tok, const char *text, int text_len, int32_t *ids, int max_ids)
void	ck_tokenizer_free (CKTokenizer *tok)
const char *	ck_tokenizer_id_to_token (const CKTokenizer *tok, int32_t id)
int	ck_tokenizer_load_binary (CKTokenizer *tok, int vocab_size, const int32_t *offsets, const char *strings, int num_merges, const int32_t *merges)
int	ck_tokenizer_load_gguf (CKTokenizer *tok, const char *path)
int	ck_tokenizer_load_json (CKTokenizer *tok, const char *path)
int	ck_tokenizer_load_merges (CKTokenizer *tok, const char *path)
int	ck_tokenizer_load_text (CKTokenizer *tok, const char *path)
int32_t	ck_tokenizer_lookup (const CKTokenizer *tok, const char *token)
void	ck_tokenizer_reset (CKTokenizer *tok)
void	ck_tokenizer_set_add_bos_eos (CKTokenizer *tok, bool add_bos, bool add_eos)
void	ck_tokenizer_set_add_space_prefix (CKTokenizer *tok, bool add_space_prefix)
void	ck_tokenizer_set_space_prefix_style (CKTokenizer *tok, CKSpacePrefixStyle style)
void	ck_tokenizer_set_special_ids (CKTokenizer *tok, int32_t unk, int32_t bos, int32_t eos, int32_t pad, int32_t mask)
void	ck_tokenizer_set_spm_mode (CKTokenizer *tok, CKSpmMode spm_mode)
void	ck_tokenizer_set_use_trie (CKTokenizer *tok, bool use_trie)
static int32_t	find_longest_match (const CKTokenizer *tok, const char *text, size_t text_len, size_t pos, size_t *match_len)
static int32_t	find_longest_match_hash (const CKTokenizer *tok, const char *text, size_t text_len, size_t pos, size_t *match_len)
static int32_t	find_longest_match_trie (const CKTokenizer *tok, const char *text, size_t text_len, size_t pos, size_t *match_len)
static int	preprocess_bpe_spaces (const char *text, int text_len, char *out, int out_max, CKSpacePrefixStyle style)

Function Documentation

ck_tokenizer_add_merge()

int ck_tokenizer_add_merge	(	CKTokenizer *	tok,
		int32_t	left_id,
		int32_t	right_id,
		int32_t	merged_id,
		int32_t	priority
	)

Add a BPE merge rule.

Parameters

tok	Tokenizer
left_id	Left token ID
right_id	Right token ID
merged_id	Merged token ID
priority	Lower = higher priority (applied first)

Returns

0 on success, -1 on error

Definition at line 560 of file tokenizer.c.

                                                                                                            {
    (void)tok; (void)left; (void)right; (void)merged; (void)priority; return 0;
}

References left, priority, and right.

ck_tokenizer_add_special_token()

int ck_tokenizer_add_special_token	(	CKTokenizer *	tok,
		const char *	name,
		int32_t	id
	)

Add special token (UNK, BOS, EOS, PAD, MASK).

Parameters

tok	Tokenizer
name	Special token name ("unk", "bos", "eos", "pad", "mask")
id	Token ID

Returns

0 on success, -1 on error

Definition at line 218 of file tokenizer.c.

                                                                                   {
    if (!tok || !name) return -1;
    if (ck_tokenizer_add_token(tok, name, id, -1e10f) != 0) return -1;
 
    TokenInfo *info = (TokenInfo *)ck_tokenizer_hash_table_lookup(tok->vocab, name);
    if (info) info->is_special = true;
 
    /* Also add to trie as special */
    if (tok->vocab_trie) {
        ck_trie_insert(tok->vocab_trie, name, id, true, 0);
    }
 
    if (strcmp(name, "<unk>") == 0 || strcmp(name, "[UNK]") == 0) tok->unk_id = id;
    else if (strcmp(name, "<s>") == 0 || strcmp(name, "<bos>") == 0 || strcmp(name, "[BOS]") == 0) tok->bos_id = id;
    else if (strcmp(name, "</s>") == 0 || strcmp(name, "<eos>") == 0 || strcmp(name, "[EOS]") == 0) tok->eos_id = id;
    else if (strcmp(name, "<pad>") == 0 || strcmp(name, "[PAD]") == 0) tok->pad_id = id;
 
    return 0;
}

References CKTokenizer::bos_id, ck_tokenizer_add_token(), ck_tokenizer_hash_table_lookup(), ck_trie_insert(), CKTokenizer::eos_id, id, CKTokenizer::pad_id, CKTokenizer::unk_id, CKTokenizer::vocab, and CKTokenizer::vocab_trie.

Referenced by main().

ck_tokenizer_add_token()

int ck_tokenizer_add_token	(	CKTokenizer *	tok,
		const char *	token,
		int32_t	id,
		float	score
	)

Add a token to vocabulary.

Parameters

tok	Tokenizer
token	Token string
id	Token ID
score	Token score (for SPM)

Returns

0 on success, -1 on error

Definition at line 162 of file tokenizer.c.

                                                                                         {
    if (!tok || !token) {
        return -1;
    }
 
    /* Ensure we have space in reverse vocab */
    if (id >= (int32_t)tok->vocab_capacity) {
        size_t new_cap = tok->vocab_capacity * 2;
        while (new_cap <= (size_t)id) {
            new_cap *= 2;
        }
        char **new_array = (char **)realloc(tok->id_to_token, new_cap * sizeof(char *));
        if (!new_array) {
            return -1;
        }
        memset(new_array + tok->vocab_capacity, 0, (new_cap - tok->vocab_capacity) * sizeof(char *));
        tok->id_to_token = new_array;
        tok->vocab_capacity = new_cap;
    }
 
    /* Check if token already exists */
    TokenInfo *existing = (TokenInfo *)ck_tokenizer_hash_table_lookup(tok->vocab, token);
    if (existing) {
        existing->id = id;
        existing->score = score;
        if (id >= (int32_t)tok->vocab_size) tok->vocab_size = id + 1;
        if (tok->id_to_token[id]) free(tok->id_to_token[id]);
        tok->id_to_token[id] = strdup(token);
        return 0;
    }
 
    /* Create new token info */
    TokenInfo *info = (TokenInfo *)malloc(sizeof(TokenInfo));
    if (!info) return -1;
    info->id = id;
    info->score = score;
    info->is_special = false;
 
    if (ck_tokenizer_hash_table_insert(tok->vocab, token, info) != 0) {
        free(info);
        return -1;
    }
 
    /* Also add to trie for fast longest-match lookups */
    if (tok->vocab_trie) {
        ck_trie_insert(tok->vocab_trie, token, id, false, 0);
    }
 
    if (id >= (int32_t)tok->vocab_size) tok->vocab_size = id + 1;
    if (tok->id_to_token[id]) free(tok->id_to_token[id]);
    tok->id_to_token[id] = strdup(token);
 
    return 0;
}

References ck_tokenizer_hash_table_insert(), ck_tokenizer_hash_table_lookup(), ck_trie_insert(), id, CKTokenizer::id_to_token, score, token, CKTokenizer::vocab, CKTokenizer::vocab_capacity, CKTokenizer::vocab_size, and CKTokenizer::vocab_trie.

Referenced by ck_tokenizer_add_special_token().

ck_tokenizer_create()

CKTokenizer * ck_tokenizer_create

(

CKTokenizerType

type

)

Definition at line 39 of file tokenizer.c.

                                                       {
    CKTokenizer *tok = (CKTokenizer *)malloc(sizeof(CKTokenizer));
    if (!tok) {
        return NULL;
    }
 
    memset(tok, 0, sizeof(*tok));
 
    /* Create hash table for vocabulary */
    tok->vocab = ck_tokenizer_hash_table_create(CK_TOKENIZER_HT_BUCKETS_LARGE);
    if (!tok->vocab) {
        free(tok);
        return NULL;
    }
 
    /* Create trie for fast lookups (1M nodes for ~50k vocab) */
    tok->vocab_trie = ck_trie_create(1000000);
    if (!tok->vocab_trie) {
        ck_tokenizer_hash_table_free(tok->vocab, true);
        free(tok);
        return NULL;
    }
 
    /* Initialize reverse vocab */
    tok->vocab_capacity = 4096;
    tok->id_to_token = (char **)calloc(tok->vocab_capacity, sizeof(char *));
    if (!tok->id_to_token) {
        ck_tokenizer_hash_table_free(tok->vocab, true);
        free(tok);
        return NULL;
    }
 
    /* Set default special tokens */
    tok->unk_id = 0;
    tok->bos_id = 1;
    tok->eos_id = 2;
    tok->pad_id = -1;
    tok->mask_id = -1;
 
    /* Initialize scores and types for SPM */
    tok->scores = NULL;
    tok->types = NULL;
 
    /* Set config */
    tok->config.type = type;
    tok->config.add_bos = false;
    tok->config.add_eos = false;
    tok->config.add_space_prefix = true;
    tok->config.unk_score = -1e10f;
    tok->config.spm_mode = CK_SPM_MODE_UNIGRAM;
 
    ck_tokenizer_mempool_init(&tok->pool, 1024 * 1024);
 
    return tok;
}

References CKTokenizerConfig::add_bos, CKTokenizerConfig::add_eos, CKTokenizerConfig::add_space_prefix, CKTokenizer::bos_id, CK_SPM_MODE_UNIGRAM, ck_tokenizer_hash_table_create(), ck_tokenizer_hash_table_free(), CK_TOKENIZER_HT_BUCKETS_LARGE, ck_tokenizer_mempool_init(), ck_trie_create(), CKTokenizer::config, CKTokenizer::eos_id, CKTokenizer::id_to_token, CKTokenizer::mask_id, CKTokenizer::pad_id, CKTokenizer::pool, CKTokenizer::scores, CKTokenizerConfig::spm_mode, CKTokenizerConfig::type, CKTokenizer::types, CKTokenizer::unk_id, CKTokenizerConfig::unk_score, CKTokenizer::vocab, CKTokenizer::vocab_capacity, and CKTokenizer::vocab_trie.

Referenced by ck_tokenizer_create_bpe(), ck_tokenizer_create_spm(), ck_tokenizer_create_wordpiece(), and main().

ck_tokenizer_decode()

int ck_tokenizer_decode	(	const CKTokenizer *	tok,
		const int32_t *	ids,
		int	num_ids,
		char *	text,
		int	max_len
	)

Decode token IDs to text.

Parameters

tok	Tokenizer
ids	Input token IDs
num_ids	Number of IDs
text	Output text buffer
max_len	Maximum text length

Returns

Number of bytes written

Definition at line 494 of file tokenizer.c.

                                                                                                          {
    if (!tok || !ids || !text || max_len <= 0) return 0;
    int len = 0;
    for (int i = 0; i < num_ids; i++) {
        int32_t id = ids[i];
        if (id < 0) continue;
        const char *token = ck_tokenizer_id_to_token(tok, id);
        if (!token) continue;
        int token_len = (int)strlen(token);
 
        /* Check for space prefix markers and convert to ASCII space */
        unsigned char c0 = (unsigned char)token[0];
        unsigned char c1 = (unsigned char)token[1];
 
        if (c0 == 0xC4 && c1 == 0xA0) {
            /* Ġ (U+0120) is 2 bytes - convert to space */
            if (len < max_len - 1) text[len++] = ' ';
            token += 2; token_len -= 2;
        } else if (c0 == 0xE2 && c1 == 0x96 && (unsigned char)token[2] == 0x81) {
            /* ▁ (U+2581) is 3 bytes - convert to space */
            if (len < max_len - 1) text[len++] = ' ';
            token += 3; token_len -= 3;
        }
 
        /* SentencePiece byte tokens should decode back to the raw byte rather
         * than printing the literal vocabulary piece "<0xXX>". */
        if (token_len == 6 && token[0] == '<' && token[1] == '0' &&
            token[2] == 'x' && token[5] == '>') {
            unsigned int byte_val = 0;
            if (sscanf(token, "<0x%02X>", &byte_val) == 1 && byte_val < 256) {
                if (len < max_len - 1) text[len++] = (char)byte_val;
                continue;
            }
        }
 
        for (int j = 0; j < token_len && len < max_len - 1; ) {
            if (j + 2 < token_len &&
                (unsigned char)token[j] == 0xE2 &&
                (unsigned char)token[j + 1] == 0x96 &&
                (unsigned char)token[j + 2] == 0x81) {
                text[len++] = ' ';
                j += 3;
                continue;
            }
            text[len++] = token[j++];
        }
    }
    text[len] = '\0';
    return len;
}

References CKTokenizer::bos_id, ck_tokenizer_id_to_token(), CKTokenizer::eos_id, ids, max_len, num_ids, CKTokenizer::pad_id, text, and token.

Referenced by main().

ck_tokenizer_detect_space_prefix_style()

CKSpacePrefixStyle ck_tokenizer_detect_space_prefix_style

(

CKTokenizer *

tok

)

Definition at line 281 of file tokenizer.c.

                                                                            {
    if (!tok) return CK_SPACE_PREFIX_GPT2;
 
    /* Already detected? */
    if (tok->config.space_prefix_detected && tok->config.space_prefix_style != CK_SPACE_PREFIX_AUTO) {
        return tok->config.space_prefix_style;
    }
 
    /* Count tokens starting with each style:
     * Ġ (U+0120) = bytes 0xC4 0xA0
     * ▁ (U+2581) = bytes 0xE2 0x96 0x81
     */
    int gpt2_count = 0;
    int spm_count = 0;
 
    for (size_t i = 0; i < tok->vocab_size; i++) {
        const char *token = tok->id_to_token[i];
        if (!token) continue;
 
        unsigned char c0 = (unsigned char)token[0];
        unsigned char c1 = (unsigned char)token[1];
 
        /* Check for Ġ (0xC4 0xA0) */
        if (c0 == 0xC4 && c1 == 0xA0) {
            gpt2_count++;
        }
        /* Check for ▁ (0xE2 0x96 0x81) */
        else if (c0 == 0xE2 && c1 == 0x96 && (unsigned char)token[2] == 0x81) {
            spm_count++;
        }
    }
 
    /* Determine style based on counts */
    CKSpacePrefixStyle detected;
    if (spm_count > gpt2_count * 2 && spm_count > 0) {
        detected = CK_SPACE_PREFIX_SPM;
    } else if (gpt2_count > 0) {
        detected = CK_SPACE_PREFIX_GPT2;
    } else {
        detected = CK_SPACE_PREFIX_ASCII;
    }
 
    tok->config.space_prefix_style = detected;
    tok->config.space_prefix_detected = true;
 
    return detected;
}

References CK_SPACE_PREFIX_ASCII, CK_SPACE_PREFIX_AUTO, CK_SPACE_PREFIX_GPT2, CK_SPACE_PREFIX_SPM, CKTokenizer::config, CKTokenizer::id_to_token, CKTokenizerConfig::space_prefix_detected, CKTokenizerConfig::space_prefix_style, token, and CKTokenizer::vocab_size.

Referenced by ck_tokenizer_encode().

ck_tokenizer_encode()

int ck_tokenizer_encode	(	const CKTokenizer *	tok,
		const char *	text,
		int	text_len,
		int32_t *	ids,
		int	max_ids
	)

Encode text to token IDs using greedy longest-match.

For BPE: applies merge rules iteratively. For WordPiece/SPM: greedy longest-match from vocabulary.

Parameters

tok	Tokenizer
text	Input text
text_len	Text length, or -1 for null-terminated
ids	Output token IDs
max_ids	Maximum IDs to write

Returns

Number of tokens written

Definition at line 436 of file tokenizer.c.

                                                                                                           {
    if (!tok || !text || !ids || max_ids <= 0) return 0;
    if (text_len < 0) text_len = (int)strlen(text);
 
    /* SentencePiece implementation lives in tokenizer_spm.c */
    if (tok->config.type == CK_TOKENIZER_SPM) {
        return ck_tokenizer_encode_spm_dispatch(tok, text, text_len, ids, max_ids);
    }
 
    /* For BPE tokenizers, convert spaces to appropriate prefix marker.
     * Auto-detect style from vocabulary if not already set. */
    char preprocessed[8192];
    const char *input = text;
    int input_len = text_len;
 
    if (tok->config.type == CK_TOKENIZER_BPE) {
        /* Get or detect space prefix style */
        CKSpacePrefixStyle style = ((CKTokenizer *)tok)->config.space_prefix_style;
        if (!((CKTokenizer *)tok)->config.space_prefix_detected) {
            style = ck_tokenizer_detect_space_prefix_style((CKTokenizer *)tok);
        }
 
        int pp_len = preprocess_bpe_spaces(text, text_len, preprocessed, sizeof(preprocessed) - 1, style);
        if (pp_len > 0) {
            preprocessed[pp_len] = '\0';
            input = preprocessed;
            input_len = pp_len;
        }
    }
 
    int out_idx = 0;
    if (tok->config.add_bos && tok->bos_id >= 0 && out_idx < max_ids) {
        ids[out_idx++] = tok->bos_id;
    }
 
    size_t pos = 0;
    while (pos < (size_t)input_len && out_idx < max_ids) {
        size_t match_len = 0;
        int32_t id = find_longest_match(tok, input, input_len, pos, &match_len);
 
        if (match_len == 0) {
            /* Emit UNK for unknown characters */
            if (tok->unk_id >= 0) ids[out_idx++] = tok->unk_id;
            pos++;
        } else {
            ids[out_idx++] = id;
            pos += match_len;
        }
    }
 
    if (tok->config.add_eos && tok->eos_id >= 0 && out_idx < max_ids) {
        ids[out_idx++] = tok->eos_id;
    }
 
    return out_idx;
}

References CKTokenizer::add_bos, CKTokenizerConfig::add_bos, CKTokenizer::add_eos, CKTokenizerConfig::add_eos, CKTokenizer::bos_id, CK_TOKENIZER_BPE, ck_tokenizer_detect_space_prefix_style(), ck_tokenizer_encode_spm_dispatch(), ck_tokenizer_lookup(), ck_tokenizer_lookup_merge(), CK_TOKENIZER_SPM, CKTokenizer::config, config, CKTokenizer::eos_id, find_longest_match(), id, ids, max_ids, CKMergeRule::merged, CKTokenizer::merges, CKTokenizer::num_merges, out_len, preprocess_bpe_spaces(), CKMergeRule::priority, style, text, text_len, CKTokenizerConfig::type, CKTokenizer::unk_id, and utf8_len().

Referenced by main(), and run_inference().

ck_tokenizer_encode_spm_dispatch()

int ck_tokenizer_encode_spm_dispatch	(	const CKTokenizer *	tok,
		const char *	text,
		int	text_len,
		int32_t *	ids,
		int	max_ids
	)

Definition at line 792 of file tokenizer_spm.c.

                                                  {
    if (!tok || !text || !ids || max_ids <= 0) return 0;
    if (text_len < 0) text_len = (int)strlen(text);
 
    int out_idx = 0;
    if (tok->config.add_bos && tok->bos_id >= 0 && out_idx < max_ids) {
        ids[out_idx++] = tok->bos_id;
    }
    if (text_len == 0) {
        if (tok->config.add_eos && tok->eos_id >= 0 && out_idx < max_ids) {
            ids[out_idx++] = tok->eos_id;
        }
        return out_idx;
    }
 
    int segment_start = 0;
    for (int pos = 0; pos < text_len && out_idx < max_ids;) {
        size_t special_len = 0;
        int32_t special_id = spm_find_special_token_at_pos(tok, text, text_len, (size_t)pos, &special_len);
        if (special_id < 0 || special_len == 0) {
            pos++;
            continue;
        }
 
        if (segment_start < pos) {
            int n = ck_tokenizer_encode_spm_plain_segment(
                tok,
                text + segment_start,
                pos - segment_start,
                ids + out_idx,
                max_ids - out_idx
            );
            if (n <= 0) return n;
            out_idx += n;
            if (out_idx >= max_ids) break;
        }
 
        ids[out_idx++] = special_id;
        pos += (int)special_len;
        segment_start = pos;
    }
 
    if (segment_start < text_len && out_idx < max_ids) {
        int n = ck_tokenizer_encode_spm_plain_segment(
            tok,
            text + segment_start,
            text_len - segment_start,
            ids + out_idx,
            max_ids - out_idx
        );
        if (n <= 0) return n;
        out_idx += n;
    }
 
    if (tok->config.add_eos && tok->eos_id >= 0 && out_idx < max_ids) {
        ids[out_idx++] = tok->eos_id;
    }
 
    return out_idx;
}

References CKTokenizerConfig::add_bos, CKTokenizerConfig::add_eos, CKTokenizer::bos_id, ck_tokenizer_encode_spm_plain_segment(), CKTokenizer::config, CKTokenizer::eos_id, ids, max_ids, spm_find_special_token_at_pos(), text, and text_len.

Referenced by ck_tokenizer_encode().

ck_tokenizer_free()

void ck_tokenizer_free

(

CKTokenizer *

tok

)

Definition at line 96 of file tokenizer.c.

                                         {
    if (!tok) return;
 
    /* Free vocabulary entries */
    if (tok->vocab) {
        ck_tokenizer_hash_table_free(tok->vocab, true);
    }
 
    /* Free trie */
    if (tok->vocab_trie) {
        ck_trie_free(tok->vocab_trie);
    }
 
    /* Free reverse vocab strings */
    if (tok->id_to_token) {
        /* Note: strings were strdup'd in add_token */
        for (size_t i = 0; i < tok->vocab_size; i++) {
            if (tok->id_to_token[i]) {
                free(tok->id_to_token[i]);
            }
        }
        free(tok->id_to_token);
    }
 
    /* Free SPM-related arrays */
    if (tok->scores) free(tok->scores);
    if (tok->types) free(tok->types);
    if (tok->byte_token_id) free(tok->byte_token_id);
 
    ck_tokenizer_mempool_free(&tok->pool);
    free(tok);
}

References CKTokenizer::byte_token_id, ck_pool_free(), ck_tokenizer_hash_table_free(), ck_tokenizer_mempool_free(), ck_trie_free(), CKTokenizer::id_to_token, CKTokenizer::merge_hash, CKTokenizer::merges, CKTokenizer::pool, CKTokenizer::scores, CKTokenizer::types, CKTokenizer::vocab, CKTokenizer::vocab_hash, CKTokenizer::vocab_size, and CKTokenizer::vocab_trie.

Referenced by main(), run_inference(), run_inference(), and run_inference().

ck_tokenizer_id_to_token()

const char * ck_tokenizer_id_to_token	(	const CKTokenizer *	tok,
		int32_t	id
	)

Definition at line 337 of file tokenizer.c.

                                                                         {
    if (!tok || id < 0 || id >= (int32_t)tok->vocab_size) return NULL;
    return tok->id_to_token[id];
}

References id, CKTokenizer::id_to_token, and CKTokenizer::vocab_size.

Referenced by ck_tokenizer_decode(), ck_tokenizer_encode_spm_impl(), main(), and run_inference().

ck_tokenizer_load_binary()

int ck_tokenizer_load_binary	(	CKTokenizer *	tok,
		int	vocab_size,
		const int32_t *	offsets,
		const char *	strings,
		int	num_merges,
		const int32_t *	merges
	)

Load vocabulary from memory-mapped binary data.

Parameters

tok	Tokenizer
vocab_size	Number of tokens
offsets	Array of offsets into strings pool
strings	String pool containing null-terminated tokens
num_merges	Number of BPE merges
merges	Merge rules as (left, right, merged) triplets

Returns

0 on success, -1 on error

Definition at line 546 of file tokenizer.c.

                                                    {
    return ck_tokenizer_load_binary_with_scores(tok, vocab_size, offsets, strings, NULL, NULL, num_merges, merges);
}

References ck_tokenizer_load_binary_with_scores(), merges, num_merges, offsets, strings, and vocab_size.

ck_tokenizer_load_gguf()

int ck_tokenizer_load_gguf	(	CKTokenizer *	tok,
		const char *	path
	)

Load vocabulary from GGUF file.

Parameters

tok	Tokenizer
path	Path to GGUF file

Returns

0 on success, -1 on error

Definition at line 556 of file tokenizer.c.

{ (void)tok; (void)path; return -1; }

ck_tokenizer_load_json()

int ck_tokenizer_load_json	(	CKTokenizer *	tok,
		const char *	path
	)

Load vocabulary from JSON file (HuggingFace format).

Parameters

tok	Tokenizer
path	Path to vocab.json or tokenizer.json

Returns

0 on success, -1 on error

Definition at line 557 of file tokenizer.c.

{ (void)tok; (void)path; return -1; }

ck_tokenizer_load_merges()

int ck_tokenizer_load_merges	(	CKTokenizer *	tok,
		const char *	path
	)

Load BPE merges from text file.

Format: token1 token2 (one merge per line)

Parameters

tok	Tokenizer
path	Path to merges.txt

Returns

0 on success, -1 on error

Definition at line 559 of file tokenizer.c.

{ (void)tok; (void)path; return -1; }

ck_tokenizer_load_text()

int ck_tokenizer_load_text	(	CKTokenizer *	tok,
		const char *	path
	)

Load vocabulary from text file (one token per line).

Format: token_string [id] [score] Lines starting with # are comments.

Parameters

tok	Tokenizer
path	Path to vocabulary file

Returns

0 on success, -1 on error

Definition at line 558 of file tokenizer.c.

{ (void)tok; (void)path; return -1; }

ck_tokenizer_lookup()

int32_t ck_tokenizer_lookup	(	const CKTokenizer *	tok,
		const char *	token
	)

Definition at line 330 of file tokenizer.c.

                                                                       {
    if (!tok || !token) return -1;
    TokenInfo *info = (TokenInfo *)ck_tokenizer_hash_table_lookup(tok->vocab, token);
    return info ? info->id : tok->unk_id;
}

References ck_tokenizer_hash_table_lookup(), token, CKTokenizer::unk_id, and CKTokenizer::vocab.

ck_tokenizer_reset()

void ck_tokenizer_reset

(

CKTokenizer *

tok

)

Definition at line 130 of file tokenizer.c.

                                          {
    if (!tok) return;
 
    ck_tokenizer_hash_table_clear(tok->vocab, true);
 
    if (tok->vocab_trie) {
        ck_trie_clear(tok->vocab_trie);
    }
 
    for (size_t i = 0; i < tok->vocab_size; i++) {
        if (tok->id_to_token[i]) {
            free(tok->id_to_token[i]);
            tok->id_to_token[i] = NULL;
        }
    }
 
    tok->vocab_size = 0;
 
    /* Reset SPM-related arrays using actual allocated sizes */
    if (tok->scores && tok->scores_size > 0) {
        memset(tok->scores, 0, tok->scores_size * sizeof(float));
    }
    if (tok->types && tok->types_size > 0) {
        memset(tok->types, 0, tok->types_size * sizeof(uint8_t));
    }
    /* Clear byte lookup table */
    if (tok->byte_token_id) {
        memset(tok->byte_token_id, -1, 256 * sizeof(int32_t));
    }
}

References CKTokenizer::byte_token_id, ck_tokenizer_hash_table_clear(), ck_trie_clear(), CKTokenizer::id_to_token, CKTokenizer::scores, CKTokenizer::scores_size, CKTokenizer::types, CKTokenizer::types_size, CKTokenizer::vocab, CKTokenizer::vocab_size, and CKTokenizer::vocab_trie.

Referenced by ck_tokenizer_load_binary_with_scores().

ck_tokenizer_set_add_bos_eos()

void ck_tokenizer_set_add_bos_eos	(	CKTokenizer *	tok,
		bool	add_bos,
		bool	add_eos
	)

Definition at line 248 of file tokenizer.c.

                                                                                {
    if (!tok) return;
    tok->config.add_bos = add_bos;
    tok->config.add_eos = add_eos;
}

References CKTokenizerConfig::add_bos, add_bos, CKTokenizerConfig::add_eos, add_eos, and CKTokenizer::config.

ck_tokenizer_set_add_space_prefix()

void ck_tokenizer_set_add_space_prefix	(	CKTokenizer *	tok,
		bool	add_space_prefix
	)

Definition at line 254 of file tokenizer.c.

                                                                                {
    if (!tok) return;
    tok->config.add_space_prefix = add_space_prefix;
}

References CKTokenizerConfig::add_space_prefix, add_space_prefix, and CKTokenizer::config.

ck_tokenizer_set_space_prefix_style()

void ck_tokenizer_set_space_prefix_style	(	CKTokenizer *	tok,
		CKSpacePrefixStyle	style
	)

Definition at line 271 of file tokenizer.c.

                                                                                     {
    if (!tok) return;
    tok->config.space_prefix_style = style;
    if (style != CK_SPACE_PREFIX_AUTO) {
        tok->config.space_prefix_detected = true;
    }
}

References CK_SPACE_PREFIX_AUTO, CKTokenizer::config, CKTokenizerConfig::space_prefix_detected, CKTokenizerConfig::space_prefix_style, and style.

ck_tokenizer_set_special_ids()

void ck_tokenizer_set_special_ids	(	CKTokenizer *	tok,
		int32_t	unk,
		int32_t	bos,
		int32_t	eos,
		int32_t	pad,
		int32_t	mask
	)

Definition at line 239 of file tokenizer.c.

                                                                                                                      {
    if (!tok) return;
    tok->unk_id = unk;
    tok->bos_id = bos;
    tok->eos_id = eos;
    tok->pad_id = pad;
    tok->mask_id = mask;
}

References bos, CKTokenizer::bos_id, eos, CKTokenizer::eos_id, mask, CKTokenizer::mask_id, pad, CKTokenizer::pad_id, unk, and CKTokenizer::unk_id.

ck_tokenizer_set_spm_mode()

void ck_tokenizer_set_spm_mode	(	CKTokenizer *	tok,
		CKSpmMode	spm_mode
	)

Definition at line 259 of file tokenizer.c.

                                                                     {
    if (!tok) return;
    tok->config.spm_mode = spm_mode;
}

References CKTokenizer::config, CKTokenizerConfig::spm_mode, and spm_mode.

ck_tokenizer_set_use_trie()

void ck_tokenizer_set_use_trie	(	CKTokenizer *	tok,
		bool	use_trie
	)

Definition at line 265 of file tokenizer.c.

                                                                {
    if (!tok) return;
    tok->config.use_trie = use_trie;
}

References CKTokenizer::config, CKTokenizerConfig::use_trie, and use_trie.

find_longest_match()

static int32_t find_longest_match ( const CKTokenizer *  tok, const char *  text, size_t  text_len, size_t  pos, size_t *  match_len  )

static

Definition at line 384 of file tokenizer.c.

                                                                                                                            {
    if (tok->config.use_trie) {
        return find_longest_match_trie(tok, text, text_len, pos, match_len);
    } else {
        return find_longest_match_hash(tok, text, text_len, pos, match_len);
    }
}

References CKTokenizer::config, find_longest_match_hash(), find_longest_match_trie(), text, text_len, and CKTokenizerConfig::use_trie.

Referenced by ck_tokenizer_encode().

find_longest_match_hash()

static int32_t find_longest_match_hash ( const CKTokenizer *  tok, const char *  text, size_t  text_len, size_t  pos, size_t *  match_len  )

static

Definition at line 354 of file tokenizer.c.

                                                                                                                                 {
    if (!tok || !text || pos >= text_len) {
        *match_len = 0;
        return tok ? tok->unk_id : -1;
    }
 
    size_t max_len = 64;
    if (pos + max_len > text_len) max_len = text_len - pos;
 
    int32_t best_id = tok->unk_id;
    size_t best_len = 0;
 
    for (size_t len = max_len; len >= 1; len--) {
        char tmp[65];
        memcpy(tmp, text + pos, len);
        tmp[len] = '\0';
 
        TokenInfo *info = (TokenInfo *)ck_tokenizer_hash_table_lookup(tok->vocab, tmp);
        if (info) {
            best_id = info->id;
            best_len = len;
            break;
        }
    }
 
    *match_len = best_len;
    return best_id;
}

References ck_tokenizer_hash_table_lookup(), max_len, text, text_len, CKTokenizer::unk_id, and CKTokenizer::vocab.

Referenced by find_longest_match().

find_longest_match_trie()

static int32_t find_longest_match_trie ( const CKTokenizer *  tok, const char *  text, size_t  text_len, size_t  pos, size_t *  match_len  )

static

Definition at line 343 of file tokenizer.c.

                                                                                                                                 {
    if (!tok || !tok->vocab_trie || !text || pos >= text_len) {
        *match_len = 0;
        return tok ? tok->unk_id : -1;
    }
 
    int32_t token_id = ck_trie_find_longest(tok->vocab_trie, text, text_len, pos, match_len);
    return token_id >= 0 ? token_id : tok->unk_id;
}

References ck_trie_find_longest(), text, text_len, CKTokenizer::unk_id, and CKTokenizer::vocab_trie.

Referenced by find_longest_match().

preprocess_bpe_spaces()

static int preprocess_bpe_spaces ( const char *  text, int  text_len, char *  out, int  out_max, CKSpacePrefixStyle  style  )

static

Definition at line 396 of file tokenizer.c.

                                                                                                                   {
    int out_len = 0;
 
    if (style == CK_SPACE_PREFIX_ASCII) {
        if (text_len > out_max) return -1;
        memcpy(out, text, (size_t)text_len);
        return text_len;
    }
 
    /* For SentencePiece, add ▁ at the start of text (unless text starts with space) */
    if (style == CK_SPACE_PREFIX_SPM && text_len > 0 && text[0] != ' ') {
        if (out_len + 3 > out_max) return -1;
        out[out_len++] = (char)0xE2;
        out[out_len++] = (char)0x96;
        out[out_len++] = (char)0x81;
    }
 
    for (int i = 0; i < text_len; i++) {
        if (text[i] == ' ') {
            if (style == CK_SPACE_PREFIX_SPM) {
                /* SentencePiece style: ▁ (3 bytes: 0xE2 0x96 0x81) */
                if (out_len + 3 > out_max) return -1;
                out[out_len++] = (char)0xE2;
                out[out_len++] = (char)0x96;
                out[out_len++] = (char)0x81;
            } else {
                /* GPT-2 style: Ġ (2 bytes: 0xC4 0xA0) */
                if (out_len + 2 > out_max) return -1;
                out[out_len++] = (char)0xC4;
                out[out_len++] = (char)0xA0;
            }
        } else {
            if (out_len + 1 > out_max) return -1;
            out[out_len++] = text[i];
        }
    }
    return out_len;
}

References CK_SPACE_PREFIX_ASCII, CK_SPACE_PREFIX_SPM, out_len, style, text, and text_len.

Referenced by ck_tokenizer_encode().