kitoken 0.11.0

Fast tokenizer for language models, supporting BPE, Unigram and WordPiece tokenization
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154

//! Decoder for the tokenizer.

use alloc::format;
use alloc::string::String;
use alloc::vec::Vec;
use core::fmt::Debug;

use hashbrown::HashMap;

use crate::{
    Configuration, InsertionPosition, SpecialToken, SpecialTokenKind, SpecialVocab, TokenId, Vocab,
};

/// Errors encountered during decoding.
#[non_exhaustive]
#[derive(Debug, Clone, thiserror::Error)]
pub enum DecodeError {
    /// A token could not be decoded.
    #[error("invalid token {0}")]
    InvalidToken(TokenId),
}

pub(crate) type DecoderMap = HashMap<TokenId, Vec<u8>>;
pub(crate) type SpecialDecoderMap = HashMap<TokenId, SpecialToken>;

/// Decoder for the tokenizer.
#[derive(Clone)]
pub(crate) struct Decoder {
    vocab:    DecoderMap,
    specials: SpecialDecoderMap,

    subword_prefix: Option<String>,

    max_token_bytes: usize,
}
impl Decoder {
    #[inline(always)]
    pub(crate) fn new(vocab: &Vocab, specials: &SpecialVocab, config: &Configuration) -> Self {
        let max_token_bytes = vocab.iter().map(|k| k.len()).max().unwrap().max(1);
        let specials = specials.iter().map(|special| (special.id, special.clone())).collect();
        let vocab = vocab.iter().map(|token| token.into()).collect();
        let subword_prefix = config.templates.iter().find_map(|template| {
            if template.position == InsertionPosition::WordContinuation {
                Some(template.content.clone())
            } else {
                None
            }
        });
        Self {
            vocab,
            specials,
            subword_prefix,
            max_token_bytes,
        }
    }

    #[inline(never)]
    pub(crate) fn decode(
        &self, tokens: &[TokenId], decode_specials: &[SpecialTokenKind],
    ) -> Result<Vec<u8>, DecodeError> {
        let extend = self.subword_prefix.as_deref().unwrap_or_default();
        let mut result = Vec::<u8>::with_capacity(
            tokens.len() * self.max_token_bytes + tokens.len() * extend.len(),
        );
        if !extend.is_empty() {
            Self::decode_with_prefix(
                &mut result,
                tokens,
                extend,
                &self.vocab,
                &self.specials,
                decode_specials,
            )?;
        } else {
            Self::decode_direct(&mut result, tokens, &self.vocab, &self.specials, decode_specials)?;
        }
        Ok(result)
    }

    #[inline(never)]
    #[cfg_attr(
        feature = "multiversion",
        multiversion::multiversion(targets(
            "x86_64+sse3+ssse3+sse4.1+sse4.2+avx+avx2+bmi2+f16c+lzcnt+popcnt",
            "x86_64+sse3+ssse3+sse4.1+sse4.2",
            "aarch64+neon",
            "wasm32+simd128",
        ))
    )]
    fn decode_direct(
        result: &mut Vec<u8>, tokens: &[TokenId], vocab: &DecoderMap, specials: &SpecialDecoderMap,
        decode_specials: &[SpecialTokenKind],
    ) -> Result<(), DecodeError> {
        for token in tokens {
            let bytes = vocab.get(token);
            if let Some(bytes) = bytes {
                result.extend(bytes);
            } else if let Some(special) = specials.get(token) {
                if decode_specials.contains(&special.kind) {
                    result.extend(special);
                }
            } else {
                return Err(DecodeError::InvalidToken(*token));
            }
        }
        Ok(())
    }

    #[inline(never)]
    #[cfg_attr(
        feature = "multiversion",
        multiversion::multiversion(targets(
            "x86_64+sse3+ssse3+sse4.1+sse4.2+avx+avx2+bmi2+f16c+lzcnt+popcnt",
            "x86_64+sse3+ssse3+sse4.1+sse4.2",
            "aarch64+neon",
            "wasm32+simd128",
        ))
    )]
    fn decode_with_prefix(
        result: &mut Vec<u8>, tokens: &[TokenId], prefix: &str, vocab: &DecoderMap,
        specials: &SpecialDecoderMap, decode_specials: &[SpecialTokenKind],
    ) -> Result<(), DecodeError> {
        for token in tokens {
            let bytes = vocab.get(token);
            if let Some(bytes) = bytes {
                if !result.is_empty() && !bytes.starts_with(prefix.as_bytes()) {
                    result.push(b' ');
                }
                result.extend(bytes);
            } else if let Some(special) = specials.get(token) {
                if !result.is_empty() {
                    result.push(b' ');
                }
                if decode_specials.contains(&special.kind) {
                    result.extend(special);
                }
            } else {
                return Err(DecodeError::InvalidToken(*token));
            }
        }
        Ok(())
    }
}
impl Debug for Decoder {
    #[inline(never)]
    fn fmt(&self, f: &mut core::fmt::Formatter) -> core::fmt::Result {
        f.debug_struct("Decoder")
            .field("vocab", &format!("DecoderMap({})", self.vocab.len()))
            .field("specials", &format!("SpecialDecoderMap({})", self.specials.len()))
            .field("subword_prefix", &self.subword_prefix)
            .field("max_token_bytes", &self.max_token_bytes)
            .finish()
    }
}