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//! This module preserves the existing hybrid behavior: `fastokens` handles encoding and
7//! `HuggingFaceTokenizer` handles decoding. Both are loaded from the same `tokenizer.json`
8//! file.
9
10use std::path::Path;
11
12use rayon::prelude::*;
13
14use super::{
15    EncodeSegment, 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    fn encode_segments(&self, segments: &[EncodeSegment<'_>]) -> Result<Encoding> {
54        let segments: Vec<fastokens::EncodeSegment<'_>> = segments
55            .iter()
56            .map(|segment| fastokens::EncodeSegment {
57                text: segment.text,
58                allow_special: segment.allow_special,
59            })
60            .collect();
61        let ids = self
62            .fast_encoder
63            .encode_segments(&segments)
64            .map_err(|e| Error::msg(format!("Fastokens segmented encode error: {e}")))?;
65        Ok(Encoding::Sp(ids))
66    }
67}
68
69impl Decoder for FastTokenizer {
70    fn decode(&self, token_ids: &[TokenIdType], skip_special_tokens: bool) -> Result<DecodeResult> {
71        self.hf_decoder.decode(token_ids, skip_special_tokens)
72    }
73}
74
75impl Tokenizer for FastTokenizer {
76    fn validate_prefix_cache(&self) -> Result<()> {
77        Ok(())
78    }
79
80    // `fast_encoder` and `hf_decoder` are loaded from the same tokenizer.json,
81    // so the HF side's vocabulary introspection applies to both.
82    fn vocab_size(&self) -> Option<usize> {
83        self.hf_decoder.vocab_size()
84    }
85
86    fn token_to_id(&self, token: &str) -> Result<Option<TokenIdType>> {
87        self.hf_decoder.token_to_id(token)
88    }
89
90    fn special_token_ids(&self) -> Result<Vec<TokenIdType>> {
91        self.hf_decoder.special_token_ids()
92    }
93
94    fn num_special_tokens_added(&self) -> Result<usize> {
95        Ok(0)
96    }
97}
98
99#[cfg(test)]
100mod tests {
101    use super::*;
102    use crate::{HuggingFaceTokenizer, TokenizerOptions};
103
104    // Minimal synthetic BPE tokenizer with no normalizer or post-processor --
105    // compatible with fastokens. Vocab covers: H,T,a,d,e,h,i,l,o,r,s,t,w + punctuation.
106    const TOKENIZER_PATH: &str = concat!(
107        env!("CARGO_MANIFEST_DIR"),
108        "/tests/data/minimal-bpe/tokenizer.json"
109    );
110    const SEGMENTED_TOKENIZER_PATH: &str = concat!(
111        env!("CARGO_MANIFEST_DIR"),
112        "/tests/data/sample-models/TinyLlama_v1.1/tokenizer.json"
113    );
114
115    #[test]
116    fn test_fast_encode_decode_roundtrip() {
117        let tokenizer = FastTokenizer::from_file(TOKENIZER_PATH).unwrap();
118        // Encode then decode: verifies both paths execute without error.
119        // With a null decoder, HF inserts spaces between tokens so exact equality
120        // is not expected here -- we just verify the operations succeed and produce
121        // non-empty results.
122        let text = "Hello, world!";
123        let encoding = tokenizer.encode(text).unwrap();
124        assert!(!encoding.token_ids().is_empty());
125        let decoded: String = tokenizer.decode(encoding.token_ids(), true).unwrap().into();
126        assert!(!decoded.is_empty());
127        // The decoded text should contain the same non-space characters
128        let enc_chars: String = text.chars().filter(|c| !c.is_whitespace()).collect();
129        let dec_chars: String = decoded.chars().filter(|c| !c.is_whitespace()).collect();
130        assert_eq!(
131            enc_chars, dec_chars,
132            "non-space characters must be preserved"
133        );
134    }
135
136    #[test]
137    fn test_fast_matches_hf_encoding() {
138        let fast = FastTokenizer::from_file(TOKENIZER_PATH).unwrap();
139        let hf = HuggingFaceTokenizer::from_file(TOKENIZER_PATH).unwrap();
140
141        for text in &["Hello, world!", "Hello", " world", "He llo"] {
142            let fast_ids = fast.encode(text).unwrap();
143            let hf_ids = hf.encode(text).unwrap();
144            assert_eq!(
145                fast_ids.token_ids(),
146                hf_ids.token_ids(),
147                "fastokens and HuggingFace must produce identical token IDs for '{text}'"
148            );
149        }
150    }
151
152    #[test]
153    fn test_fast_batch_encode() {
154        let tokenizer = FastTokenizer::from_file(TOKENIZER_PATH).unwrap();
155        let inputs = &["Hello", " world", "Hello, world!"];
156        let encodings = tokenizer.encode_batch(inputs).unwrap();
157        assert_eq!(encodings.len(), inputs.len());
158        for (enc, input) in encodings.iter().zip(inputs.iter()) {
159            assert!(
160                !enc.token_ids().is_empty(),
161                "encoding for '{input}' must be non-empty"
162            );
163        }
164    }
165
166    #[test]
167    fn test_fast_segmented_encoding_preserves_trust_boundaries() {
168        let tokenizer = FastTokenizer::from_file(SEGMENTED_TOKENIZER_PATH).unwrap();
169        let upstream =
170            fastokens::Tokenizer::from_file(std::path::Path::new(SEGMENTED_TOKENIZER_PATH))
171                .unwrap();
172        let marker = "<s>";
173
174        let trusted = tokenizer
175            .encode_segments(&[EncodeSegment::control(marker)])
176            .unwrap();
177        assert_eq!(
178            trusted.token_ids(),
179            &[upstream.token_to_id(marker).unwrap()],
180            "trusted renderer output must recognize the control token"
181        );
182
183        let ordinary = tokenizer
184            .encode_segments(&[EncodeSegment::ordinary(marker)])
185            .unwrap();
186        assert_ne!(
187            ordinary.token_ids(),
188            trusted.token_ids(),
189            "untrusted content must encode the control-token spelling as ordinary text"
190        );
191
192        let segments = [
193            EncodeSegment::ordinary("hello "),
194            EncodeSegment::control(marker),
195            EncodeSegment::ordinary(marker),
196        ];
197        let upstream_segments = [
198            fastokens::EncodeSegment::ordinary("hello "),
199            fastokens::EncodeSegment::special(marker),
200            fastokens::EncodeSegment::ordinary(marker),
201        ];
202        let actual = tokenizer.encode_segments(&segments).unwrap();
203        let expected = upstream.encode_segments(&upstream_segments).unwrap();
204        assert_eq!(actual.token_ids(), expected);
205
206        assert!(
207            tokenizer
208                .encode_segments(&[])
209                .unwrap()
210                .token_ids()
211                .is_empty()
212        );
213    }
214
215    #[test]
216    fn test_fast_with_decode_stream() {
217        use crate::Tokenizer as TokenizerWrapper;
218        use std::sync::Arc;
219
220        let tokenizer = Arc::new(FastTokenizer::from_file(TOKENIZER_PATH).unwrap());
221        let wrapper = TokenizerWrapper::from(tokenizer);
222
223        // Encode a prompt and a continuation, then step through the decode stream
224        let prompt_ids = wrapper.encode("Hello").unwrap().token_ids().to_vec();
225        let continuation = ", world!";
226        let cont_ids = wrapper.encode(continuation).unwrap().token_ids().to_vec();
227
228        let mut stream = wrapper.decode_stream(&prompt_ids, true);
229        // Accumulate incremental chunks from decode_stream
230        let mut accumulated = String::new();
231        for id in &cont_ids {
232            if let Some(chunk) = stream.step(*id).unwrap() {
233                accumulated.push_str(&chunk);
234            }
235        }
236
237        // DecodeStream uses prompt tokens as context, so the expected text is
238        // decode(prompt + continuation) minus decode(prompt) -- not a bare
239        // decode(continuation) which lacks the surrounding context.
240        let mut all_ids = prompt_ids.clone();
241        all_ids.extend_from_slice(&cont_ids);
242        let full_text: String = wrapper.decode(&all_ids, true).unwrap().into();
243        let prompt_text: String = wrapper.decode(&prompt_ids, true).unwrap().into();
244        let expected = &full_text[prompt_text.len()..];
245        assert_eq!(
246            accumulated, expected,
247            "streamed chunks must equal context-aware decoded continuation"
248        );
249    }
250
251    #[test]
252    fn vocabulary_metadata_forwards_to_hf_decoder() {
253        let fast = FastTokenizer::from_file(TOKENIZER_PATH).unwrap();
254        let hf = HuggingFaceTokenizer::from_file(TOKENIZER_PATH).unwrap();
255        assert_eq!(fast.vocab_size(), hf.vocab_size());
256        assert_eq!(
257            fast.token_to_id("Hello").unwrap(),
258            hf.token_to_id("Hello").unwrap()
259        );
260        assert_eq!(
261            fast.special_token_ids().unwrap(),
262            hf.special_token_ids().unwrap()
263        );
264    }
265
266    #[test]
267    fn special_token_accounting_matches_fast_encoder() {
268        let fast = FastTokenizer::from_file(SEGMENTED_TOKENIZER_PATH).unwrap();
269        let upstream =
270            fastokens::Tokenizer::from_file(std::path::Path::new(SEGMENTED_TOKENIZER_PATH))
271                .unwrap();
272        let hf = HuggingFaceTokenizer::from_file(SEGMENTED_TOKENIZER_PATH).unwrap();
273
274        assert_eq!(hf.num_special_tokens_added().unwrap(), 1);
275        assert_eq!(fast.num_special_tokens_added().unwrap(), 0);
276        let hf_with_special_tokens = hf.with_options(TokenizerOptions {
277            add_special_tokens: true,
278        });
279
280        for text in ["hello", "hello there"] {
281            let fast_ids = fast.encode(text).unwrap();
282            assert_eq!(
283                fast_ids.token_ids(),
284                upstream.encode(text).unwrap(),
285                "FastTokenizer must match the encoder that omits the HF post-processor"
286            );
287            assert_eq!(
288                hf_with_special_tokens
289                    .encode(text)
290                    .unwrap()
291                    .token_ids()
292                    .len(),
293                fast_ids.token_ids().len() + 1,
294                "the HF post-processor must add the BOS token FastTokenizer omits"
295            );
296        }
297    }
298}
dynamo_tokenizers/fastokens.rs

dynamo_tokenizers/
fastokens.rs
