llm-tokenizer 1.2.0

LLM tokenizer library with caching and chat template support
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

use std::{
    collections::hash_map::DefaultHasher,
    hash::{Hash, Hasher},
};

use anyhow::Result;

use crate::chat_template::{ChatTemplateContentFormat, ChatTemplateParams};

/// Type alias for token IDs
pub type TokenIdType = u32;

/// Core encoding trait - separate from decoding for modularity
pub trait Encoder: Send + Sync {
    fn encode(&self, input: &str, add_special_tokens: bool) -> Result<Encoding>;
    fn encode_batch(&self, inputs: &[&str], add_special_tokens: bool) -> Result<Vec<Encoding>>;
}

/// Core decoding trait - can be implemented independently
pub trait Decoder: Send + Sync {
    fn decode(&self, token_ids: &[TokenIdType], skip_special_tokens: bool) -> Result<String>;
}

/// Combined tokenizer trait
pub trait Tokenizer: Encoder + Decoder {
    fn vocab_size(&self) -> usize;
    fn get_special_tokens(&self) -> &SpecialTokens;
    fn token_to_id(&self, token: &str) -> Option<TokenIdType>;
    fn id_to_token(&self, id: TokenIdType) -> Option<String>;

    /// Enable downcasting to concrete types
    fn as_any(&self) -> &dyn std::any::Any;

    /// Apply chat template to messages. Default returns an error for tokenizers without template support.
    fn apply_chat_template(
        &self,
        _messages: &[serde_json::Value],
        _params: ChatTemplateParams,
    ) -> Result<String> {
        Err(anyhow::anyhow!(
            "Chat template not supported by this tokenizer"
        ))
    }

    /// Get the content format expected by the chat template.
    fn chat_template_content_format(&self) -> ChatTemplateContentFormat {
        ChatTemplateContentFormat::default()
    }

    /// Set or override the chat template.
    ///
    /// Returns an error if the template fails to parse or the tokenizer
    /// does not support chat templates.
    fn set_chat_template(&mut self, _template: String) -> Result<()> {
        Err(anyhow::anyhow!(
            "set_chat_template is not supported by this tokenizer"
        ))
    }
}

/// Contains the results of tokenizing text: token IDs, string tokens, and their spans
#[derive(Debug, Clone)]
pub enum Encoding {
    /// Hugging Face
    Hf(Box<tokenizers::tokenizer::Encoding>),
    /// Plain token ID vector
    Plain(Vec<TokenIdType>),
    /// Tiktoken (for GPT models) - now uses u32 in tiktoken-rs 0.7.0
    Tiktoken(Vec<TokenIdType>),
}

impl Encoding {
    /// Returns a reference to token IDs - zero-copy operation
    #[inline]
    pub fn token_ids(&self) -> &[TokenIdType] {
        match self {
            Encoding::Hf(inner) => inner.get_ids(),
            Encoding::Plain(inner) => inner,
            Encoding::Tiktoken(inner) => inner,
        }
    }

    /// Get a hash of the token IDs for caching purposes
    pub fn get_hash(&self) -> u64 {
        let mut hasher = DefaultHasher::new();
        self.hash(&mut hasher);
        hasher.finish()
    }
}

/// Hash implementation for Encoding
impl Hash for Encoding {
    fn hash<H: Hasher>(&self, state: &mut H) {
        match self {
            Encoding::Hf(inner) => inner.get_ids().hash(state),
            Encoding::Plain(inner) => inner.hash(state),
            Encoding::Tiktoken(inner) => inner.hash(state),
        }
    }
}

#[derive(Debug, Clone, Default)]
pub struct SpecialTokens {
    pub bos_token: Option<String>,
    pub eos_token: Option<String>,
    pub unk_token: Option<String>,
    pub sep_token: Option<String>,
    pub pad_token: Option<String>,
    pub cls_token: Option<String>,
    pub mask_token: Option<String>,
    pub additional_special_tokens: Vec<String>,
}