dynamo_tokenizers/
fastokens.rs

1// SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2// SPDX-License-Identifier: Apache-2.0
3
4//! Fastokens backend using the `fastokens` crate for high-performance BPE encoding.
5//!
6//! `fastokens` only supports encoding, so this module provides a hybrid tokenizer that
7//! uses `fastokens` for encoding and falls back to `HuggingFaceTokenizer` for decoding.
8//! Both are loaded from the same `tokenizer.json` file.
9
10use std::path::Path;
11
12use rayon::prelude::*;
13
14use super::{
15    Encoding, Error, Result, TokenIdType,
16    hf::HuggingFaceTokenizer,
17    traits::{DecodeResult, Decoder, Encoder, Tokenizer},
18};
19
20/// Hybrid tokenizer: fast BPE encoding via `fastokens`, decoding via HuggingFace.
21///
22/// Both backends are loaded from the same `tokenizer.json` file.
23pub struct FastTokenizer {
24    fast_encoder: fastokens::Tokenizer,
25    hf_decoder: HuggingFaceTokenizer,
26}
27
28impl FastTokenizer {
29    pub fn from_file(path: &str) -> Result<Self> {
30        let fast_encoder = fastokens::Tokenizer::from_file(Path::new(path))
31            .map_err(|e| Error::msg(format!("Error loading fastokens tokenizer: {e}")))?;
32        let hf_decoder = HuggingFaceTokenizer::from_file(path)?;
33        Ok(Self {
34            fast_encoder,
35            hf_decoder,
36        })
37    }
38}
39
40impl Encoder for FastTokenizer {
41    fn encode(&self, input: &str) -> Result<Encoding> {
42        let ids = self
43            .fast_encoder
44            .encode(input)
45            .map_err(|e| Error::msg(format!("Fastokens encode error: {e}")))?;
46        Ok(Encoding::Sp(ids))
47    }
48
49    fn encode_batch(&self, inputs: &[&str]) -> Result<Vec<Encoding>> {
50        inputs.par_iter().map(|input| self.encode(input)).collect()
51    }
52}
53
54impl Decoder for FastTokenizer {
55    fn decode(&self, token_ids: &[TokenIdType], skip_special_tokens: bool) -> Result<DecodeResult> {
56        self.hf_decoder.decode(token_ids, skip_special_tokens)
57    }
58}
59
60impl Tokenizer for FastTokenizer {
61    fn validate_prefix_cache(&self) -> Result<()> {
62        Ok(())
63    }
64}
65
66#[cfg(test)]
67mod tests {
68    use super::*;
69    use crate::HuggingFaceTokenizer;
70
71    // Minimal synthetic BPE tokenizer with no normalizer or post-processor --
72    // compatible with fastokens. Vocab covers: H,T,a,d,e,h,i,l,o,r,s,t,w + punctuation.
73    const TOKENIZER_PATH: &str = concat!(
74        env!("CARGO_MANIFEST_DIR"),
75        "/tests/data/minimal-bpe/tokenizer.json"
76    );
77
78    #[test]
79    fn test_fast_encode_decode_roundtrip() {
80        let tokenizer = FastTokenizer::from_file(TOKENIZER_PATH).unwrap();
81        // Encode then decode: verifies both paths execute without error.
82        // With a null decoder, HF inserts spaces between tokens so exact equality
83        // is not expected here -- we just verify the operations succeed and produce
84        // non-empty results.
85        let text = "Hello, world!";
86        let encoding = tokenizer.encode(text).unwrap();
87        assert!(!encoding.token_ids().is_empty());
88        let decoded: String = tokenizer.decode(encoding.token_ids(), true).unwrap().into();
89        assert!(!decoded.is_empty());
90        // The decoded text should contain the same non-space characters
91        let enc_chars: String = text.chars().filter(|c| !c.is_whitespace()).collect();
92        let dec_chars: String = decoded.chars().filter(|c| !c.is_whitespace()).collect();
93        assert_eq!(
94            enc_chars, dec_chars,
95            "non-space characters must be preserved"
96        );
97    }
98
99    #[test]
100    fn test_fast_matches_hf_encoding() {
101        let fast = FastTokenizer::from_file(TOKENIZER_PATH).unwrap();
102        let hf = HuggingFaceTokenizer::from_file(TOKENIZER_PATH).unwrap();
103
104        for text in &["Hello, world!", "Hello", " world", "He llo"] {
105            let fast_ids = fast.encode(text).unwrap();
106            let hf_ids = hf.encode(text).unwrap();
107            assert_eq!(
108                fast_ids.token_ids(),
109                hf_ids.token_ids(),
110                "fastokens and HuggingFace must produce identical token IDs for '{text}'"
111            );
112        }
113    }
114
115    #[test]
116    fn test_fast_batch_encode() {
117        let tokenizer = FastTokenizer::from_file(TOKENIZER_PATH).unwrap();
118        let inputs = &["Hello", " world", "Hello, world!"];
119        let encodings = tokenizer.encode_batch(inputs).unwrap();
120        assert_eq!(encodings.len(), inputs.len());
121        for (enc, input) in encodings.iter().zip(inputs.iter()) {
122            assert!(
123                !enc.token_ids().is_empty(),
124                "encoding for '{input}' must be non-empty"
125            );
126        }
127    }
128
129    #[test]
130    fn test_fast_with_decode_stream() {
131        use crate::Tokenizer as TokenizerWrapper;
132        use std::sync::Arc;
133
134        let tokenizer = Arc::new(FastTokenizer::from_file(TOKENIZER_PATH).unwrap());
135        let wrapper = TokenizerWrapper::from(tokenizer);
136
137        // Encode a prompt and a continuation, then step through the decode stream
138        let prompt_ids = wrapper.encode("Hello").unwrap().token_ids().to_vec();
139        let continuation = ", world!";
140        let cont_ids = wrapper.encode(continuation).unwrap().token_ids().to_vec();
141
142        let mut stream = wrapper.decode_stream(&prompt_ids, true);
143        // Accumulate incremental chunks from decode_stream
144        let mut accumulated = String::new();
145        for id in &cont_ids {
146            if let Some(chunk) = stream.step(*id).unwrap() {
147                accumulated.push_str(&chunk);
148            }
149        }
150
151        // DecodeStream uses prompt tokens as context, so the expected text is
152        // decode(prompt + continuation) minus decode(prompt) -- not a bare
153        // decode(continuation) which lacks the surrounding context.
154        let mut all_ids = prompt_ids.clone();
155        all_ids.extend_from_slice(&cont_ids);
156        let full_text: String = wrapper.decode(&all_ids, true).unwrap().into();
157        let prompt_text: String = wrapper.decode(&prompt_ids, true).unwrap().into();
158        let expected = &full_text[prompt_text.len()..];
159        assert_eq!(
160            accumulated, expected,
161            "streamed chunks must equal context-aware decoded continuation"
162        );
163    }
164}
dynamo_tokenizers/fastokens.rs

dynamo_tokenizers/
fastokens.rs
