dynamo-tokenizers 4.3.0

Standalone HuggingFace, tiktoken, fastokens, and Baseten tokenizer implementations for LLM inference serving.
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
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

//! Mistral Tekken tokenizer (`tekken.json`).
//!
//! Tekken is a tiktoken-style byte-level BPE. The file stores the BPE ranks,
//! the split regex, and the special tokens. Special tokens own ids
//! `0..num_special_tokens`; an ordinary BPE token with rank `r` has id
//! `r + num_special_tokens`. This mirrors mistral-common's `Tekkenizer`.
//!
//! Plain `encode` never recognizes special-token text, matching
//! mistral-common, which encodes message text as ordinary BPE and inserts
//! structural tokens by id. Chat prompts must therefore arrive as segments
//! from a renderer that marks its control tokens ([`EncodeSegment::control`]);
//! a text-only Jinja prompt would encode `<s>[INST]` as plain text.

use std::path::Path;

use aho_corasick::{AhoCorasick, MatchKind};
use anyhow::Context as _;
use base64::Engine as _;
use fancy_regex::Regex;
use rayon::prelude::*;
use rustc_hash::FxHashMap;
use serde::Deserialize;
use tiktoken_rs::{CoreBPE, byte_pair_split};

use super::{
    EncodeSegment, Encoding, Error, Result, TokenIdType, TokenizerOptions,
    traits::{DecodeResult, Decoder, Encoder, Tokenizer},
};

/// Special tokens kept by `decode(skip_special_tokens = true)`.
///
/// Tool-call and reasoning parsers need these markers in the decoded text.
/// Upstream vLLM's `MistralTokenizer.convert_ids_to_tokens` keeps the same set
/// when skipping special tokens.
const PARSER_MARKERS: [&str; 4] = ["[TOOL_CALLS]", "[ARGS]", "[THINK]", "[/THINK]"];

/// Released tekken files reserve 1,000 special tokens. The bound keeps a
/// malformed file from allocating billions of `<SPECIAL_n>` fillers.
const MAX_SPECIAL_TOKENS: usize = 1 << 16;

/// Byte-level BPE needs every single byte as a token.
const BYTE_TOKENS: usize = 256;

#[derive(Deserialize)]
struct TekkenFile {
    config: TekkenConfig,
    vocab: Vec<TekkenTokenInfo>,
    special_tokens: Option<Vec<TekkenSpecialToken>>,
    model_settings_builder: Option<serde_json::Value>,
}

#[derive(Deserialize)]
struct TekkenConfig {
    pattern: String,
    default_vocab_size: usize,
    default_num_special_tokens: usize,
    version: String,
}

#[derive(Deserialize)]
struct TekkenTokenInfo {
    rank: usize,
    token_bytes: String,
}

#[derive(Deserialize)]
struct TekkenSpecialToken {
    rank: usize,
    token_str: String,
}

pub struct TekkenTokenizer {
    bpe: CoreBPE,
    pattern: Regex,
    special_token_matcher: AhoCorasick,
    ordinary_token_ids: FxHashMap<Vec<u8>, TokenIdType>,
    num_special_tokens: u32,
    vocab_size: u32,
    special_token_ids: FxHashMap<String, TokenIdType>,
    kept_on_skip: Vec<TokenIdType>,
    version: String,
    model_settings_builder: Option<serde_json::Value>,
}

impl TekkenTokenizer {
    /// Load a Mistral `tekken.json` tokenizer file.
    pub fn from_file(path: &str) -> Result<Self> {
        let file: TekkenFile = {
            let contents = std::fs::read(path)
                .with_context(|| format!("Failed to read tekken file '{path}'"))?;
            serde_json::from_slice(&contents)
                .with_context(|| format!("Failed to parse tekken file '{path}'"))?
        };
        Self::from_parts(file)
    }

    fn from_parts(file: TekkenFile) -> Result<Self> {
        let TekkenConfig {
            pattern,
            default_vocab_size,
            default_num_special_tokens,
            version,
        } = file.config;
        let vocab_size = u32::try_from(default_vocab_size)
            .context("tekken default_vocab_size does not fit a token id")?;
        let num_special_tokens = u32::try_from(default_num_special_tokens)
            .context("tekken default_num_special_tokens does not fit a token id")?;
        if default_num_special_tokens > MAX_SPECIAL_TOKENS {
            anyhow::bail!(
                "tekken reserves {default_num_special_tokens} special tokens, more than the supported {MAX_SPECIAL_TOKENS}"
            );
        }

        let Some(specials) = file.special_tokens else {
            anyhow::bail!(
                "tekken {version} file has no special_tokens; only files that list them are supported"
            );
        };
        if specials.len() > default_num_special_tokens {
            anyhow::bail!(
                "tekken file lists {} special tokens but reserves only {default_num_special_tokens}",
                specials.len()
            );
        }
        let Some(inner_vocab_size) = default_vocab_size.checked_sub(default_num_special_tokens)
        else {
            anyhow::bail!("tekken default_vocab_size is smaller than its special tokens");
        };
        // Byte-pair merging requires every byte to have a token.
        if inner_vocab_size < BYTE_TOKENS {
            anyhow::bail!(
                "tekken keeps {inner_vocab_size} ordinary tokens, fewer than the {BYTE_TOKENS} single bytes"
            );
        }
        if file.vocab.len() < inner_vocab_size {
            anyhow::bail!(
                "tekken vocab has {} entries, fewer than the {inner_vocab_size} the config requires",
                file.vocab.len()
            );
        }

        // Ranks the file leaves unlisted are filled the way mistral-common does.
        let mut special_tokens: Vec<String> = (0..default_num_special_tokens)
            .map(|id| format!("<SPECIAL_{id}>"))
            .collect();
        for (index, token) in specials.into_iter().enumerate() {
            if token.rank != index {
                anyhow::bail!(
                    "tekken special token '{}' has rank {}, expected {index}",
                    token.token_str,
                    token.rank
                );
            }
            // An empty special token would make CoreBPE's special-token regex
            // match the empty string and loop forever.
            if token.token_str.is_empty() {
                anyhow::bail!("tekken special token {index} is empty");
            }
            special_tokens[index] = token.token_str;
        }

        let mut special_token_ids = FxHashMap::default();
        for (id, token) in (0..num_special_tokens).zip(special_tokens) {
            if let Some(previous) = special_token_ids.insert(token, id) {
                anyhow::bail!("tekken special token ids {previous} and {id} share one string");
            }
        }

        let engine = base64::engine::general_purpose::STANDARD;
        let mut ordinary_token_ids = FxHashMap::default();
        for (rank, token) in (0..).zip(file.vocab.into_iter().take(inner_vocab_size)) {
            if token.rank != rank as usize {
                anyhow::bail!("tekken vocab entry {rank} has rank {}", token.rank);
            }
            let bytes = engine
                .decode(&token.token_bytes)
                .with_context(|| format!("Invalid base64 in tekken vocab entry {rank}"))?;
            // mistral-common requires ranks 0..256 to be the single bytes in order.
            if (rank as usize) < BYTE_TOKENS && bytes != [rank as u8] {
                anyhow::bail!("tekken vocab entry {rank} is not the single byte {rank}");
            }
            // Shifting every rank preserves merge priority and emits tekken ids.
            if ordinary_token_ids
                .insert(bytes, num_special_tokens + rank)
                .is_some()
            {
                anyhow::bail!("tekken vocab entry {rank} duplicates an earlier token");
            }
        }

        let pattern = Regex::new(&pattern).context("Error creating tekken split regex")?;
        // CoreBPE's encoder unwraps regex failures. Keep the same ranks and
        // pattern for fallible encoding, and use CoreBPE only for decoding.
        let bpe = CoreBPE::new(
            ordinary_token_ids.clone(),
            special_token_ids.clone(),
            pattern.as_str(),
        )
        .map_err(|err| Error::msg(format!("Error creating tekken BPE: {err}")))?;
        let special_token_matcher = AhoCorasick::builder()
            .match_kind(MatchKind::LeftmostFirst)
            .build(special_token_ids.keys())
            .context("Error creating tekken special-token matcher")?;

        let kept_on_skip = PARSER_MARKERS
            .iter()
            .filter_map(|marker| special_token_ids.get(*marker).copied())
            .collect();

        Ok(Self {
            bpe,
            pattern,
            special_token_matcher,
            ordinary_token_ids,
            num_special_tokens,
            vocab_size,
            special_token_ids,
            kept_on_skip,
            version,
            model_settings_builder: file.model_settings_builder,
        })
    }

    /// Whether `path` names a Mistral tekken tokenizer file (`tekken*.json`,
    /// such as `tekken.json` or mistral-common's versioned `tekken_240911.json`).
    pub fn is_tekken_file(path: &Path) -> bool {
        path.file_name()
            .and_then(std::ffi::OsStr::to_str)
            .is_some_and(|name| name.starts_with("tekken") && name.ends_with(".json"))
    }

    /// The tokenizer format version, such as `v15`.
    pub fn version(&self) -> &str {
        &self.version
    }

    /// The numeric part of [`Self::version`], such as `15` for `v15`.
    pub fn version_number(&self) -> Option<u32> {
        self.version.strip_prefix('v')?.parse().ok()
    }

    /// The raw `model_settings_builder` section (v15+), which sets the
    /// allowed values and defaults of `[MODEL_SETTINGS]` prompt fields.
    pub fn model_settings_builder(&self) -> Option<&serde_json::Value> {
        self.model_settings_builder.as_ref()
    }

    /// Reject ids outside the vocabulary before decoding.
    ///
    /// `CoreBPE::_decode_native_and_split` panics on an unknown id. Every id
    /// below `vocab_size` is known: special ids `0..N` are filled densely and
    /// ordinary ranks are contiguous and unique, as `from_parts` checks.
    fn validate_ids(&self, token_ids: &[TokenIdType]) -> Result<()> {
        match token_ids.iter().find(|&&id| id >= self.vocab_size) {
            Some(id) => anyhow::bail!(
                "token id {id} is outside the tekken vocabulary of {}",
                self.vocab_size
            ),
            None => Ok(()),
        }
    }

    fn encode_ordinary_into(&self, input: &str, token_ids: &mut Vec<u32>) -> Result<()> {
        for piece in self.pattern.find_iter(input) {
            let piece = piece.context("Failed to match tekken split regex")?;
            let bytes = piece.as_str().as_bytes();
            if let Some(&id) = self.ordinary_token_ids.get(bytes) {
                token_ids.push(id);
            } else {
                anyhow::ensure!(
                    !bytes.is_empty(),
                    "tekken split regex matched an empty string"
                );
                // Every one-byte token exists, so a missing whole piece has at
                // least two bytes. Use the same merge primitive as CoreBPE.
                token_ids.extend(
                    byte_pair_split(bytes, &self.ordinary_token_ids)
                        .into_iter()
                        .map(|piece| self.ordinary_token_ids[piece]),
                );
            }
        }
        Ok(())
    }

    fn encode_special_into(&self, input: &str, token_ids: &mut Vec<u32>) -> Result<()> {
        let mut end = 0;
        for matched in self.special_token_matcher.find_iter(input) {
            self.encode_ordinary_into(&input[end..matched.start()], token_ids)?;
            let id = self.special_token_ids[&input[matched.start()..matched.end()]];
            token_ids.push(id);
            end = matched.end();
        }
        self.encode_ordinary_into(&input[end..], token_ids)
    }
}

impl Encoder for TekkenTokenizer {
    fn encode(&self, input: &str) -> Result<Encoding> {
        let mut token_ids = Vec::new();
        self.encode_ordinary_into(input, &mut token_ids)?;
        Ok(Encoding::Sp(token_ids))
    }

    fn encode_batch(&self, inputs: &[&str]) -> Result<Vec<Encoding>> {
        inputs.par_iter().map(|input| self.encode(input)).collect()
    }

    fn encode_segments(&self, segments: &[EncodeSegment<'_>]) -> Result<Encoding> {
        let mut token_ids = Vec::new();
        for segment in segments {
            if !segment.allow_special {
                self.encode_ordinary_into(segment.text, &mut token_ids)?;
            } else if let Some(&id) = self.special_token_ids.get(segment.text) {
                // Renderers emit one control segment per marker.
                token_ids.push(id);
            } else {
                self.encode_special_into(segment.text, &mut token_ids)?;
            }
        }
        Ok(Encoding::Sp(token_ids))
    }
}

impl Decoder for TekkenTokenizer {
    fn decode(&self, token_ids: &[TokenIdType], skip_special_tokens: bool) -> Result<DecodeResult> {
        self.validate_ids(token_ids)?;
        let ids: Vec<u32> = if skip_special_tokens {
            token_ids
                .iter()
                .filter(|&&id| id >= self.num_special_tokens || self.kept_on_skip.contains(&id))
                .copied()
                .collect()
        } else {
            token_ids.to_vec()
        };

        // Same strategy as the tiktoken backend: strict UTF-8 first, then a
        // lossy fallback whose trailing U+FFFD marks an incomplete byte run.
        let bytes: Vec<u8> = self.bpe._decode_native_and_split(ids).flatten().collect();
        match String::from_utf8(bytes) {
            Ok(text) => Ok(DecodeResult::Complete(text)),
            Err(e) => {
                let text = String::from_utf8_lossy(e.as_bytes()).into_owned();
                Ok(DecodeResult::from_decoded(text))
            }
        }
    }
}

impl Tokenizer for TekkenTokenizer {
    // `validate_prefix_cache` keeps its default error: the prefix cache splits
    // text at special-token strings, but `encode` treats them as ordinary text.

    /// `add_special_tokens` is not applied: `encode` never adds `<s>`, which
    /// the Mistral renderer emits itself, as mistral-common does.
    fn with_options(self, _options: TokenizerOptions) -> Self {
        self
    }

    fn vocab_size(&self) -> Option<usize> {
        Some(self.vocab_size as usize)
    }

    fn token_to_id(&self, token: &str) -> Result<Option<TokenIdType>> {
        if let Some(&id) = self.special_token_ids.get(token) {
            return Ok(Some(id));
        }
        // A string that BPE encodes to exactly one id is that token.
        Ok(match self.encode(token)?.token_ids() {
            &[id] => Some(id),
            _ => None,
        })
    }

    fn special_token_ids(&self) -> Result<Vec<TokenIdType>> {
        Ok((0..self.num_special_tokens).collect())
    }

    // Tekken's `encode` never adds BOS/EOS; renderers emit `<s>` explicitly.
    fn num_special_tokens_added(&self) -> Result<usize> {
        Ok(0)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::DecodeStream;
    use std::sync::Arc;

    const FIXTURE: &str = concat!(
        env!("CARGO_MANIFEST_DIR"),
        "/tests/data/sample-models/mock-tekken/tekken.json"
    );
    const EXPECTED_IDS: &str = concat!(
        env!("CARGO_MANIFEST_DIR"),
        "/tests/data/sample-models/mock-tekken/expected_ids.json"
    );

    #[derive(Deserialize)]
    struct ExpectedIds {
        text: String,
        ids: Vec<TokenIdType>,
    }

    fn id_of(tokenizer: &TekkenTokenizer, token: &str) -> Option<TokenIdType> {
        tokenizer.token_to_id(token).unwrap()
    }

    fn fixture() -> TekkenTokenizer {
        TekkenTokenizer::from_file(FIXTURE).unwrap()
    }

    #[test]
    fn encode_matches_mistral_common() {
        let tokenizer = fixture();
        let expected: Vec<ExpectedIds> =
            serde_json::from_str(&std::fs::read_to_string(EXPECTED_IDS).unwrap()).unwrap();
        assert!(!expected.is_empty());
        for case in expected {
            let ids = tokenizer.encode(&case.text).unwrap();
            assert_eq!(ids.token_ids(), case.ids.as_slice(), "text {:?}", case.text);
            let decoded: String = tokenizer.decode(&case.ids, false).unwrap().into();
            assert_eq!(decoded, case.text);
        }
    }

    #[test]
    fn metadata_comes_from_the_file() {
        let tokenizer = fixture();
        assert_eq!(tokenizer.version(), "v15");
        assert_eq!(tokenizer.version_number(), Some(15));
        assert_eq!(tokenizer.vocab_size(), Some(1040));
        assert_eq!(id_of(&tokenizer, "<s>"), Some(1));
        assert_eq!(id_of(&tokenizer, "</s>"), Some(2));
        assert_eq!(id_of(&tokenizer, "[TOOL_CALLS]"), Some(9));
        assert_eq!(id_of(&tokenizer, "[THINK]"), Some(34));
        // Ranks the file leaves unnamed become numbered fillers.
        assert_eq!(id_of(&tokenizer, "<SPECIAL_39>"), Some(39));
        assert_eq!(tokenizer.special_token_ids().unwrap().len(), 40);
        assert_eq!(tokenizer.token_to_id("[ARGS]").unwrap(), Some(32));
        assert_eq!(
            tokenizer.token_to_id("a").unwrap(),
            Some(40 + u32::from(b'a'))
        );
        assert_eq!(tokenizer.token_to_id("not-a-token-xyz").unwrap(), None);
        let builder = tokenizer.model_settings_builder().unwrap();
        assert_eq!(builder["reasoning_effort"]["default"], "none");
    }

    #[test]
    fn special_token_text_is_ordinary_unless_marked_control() {
        let tokenizer = fixture();
        let inst = id_of(&tokenizer, "[INST]").unwrap();

        let plain = tokenizer.encode("[INST]").unwrap();
        assert!(!plain.token_ids().contains(&inst));

        let segments = tokenizer
            .encode_segments(&[
                EncodeSegment::control("<s>[INST]"),
                EncodeSegment::ordinary("hi [INST]"),
                EncodeSegment::control("[/INST]"),
            ])
            .unwrap();
        let ids = segments.token_ids();
        assert_eq!(ids[..2], [1, inst]);
        assert_eq!(*ids.last().unwrap(), 4);
        assert_eq!(ids.iter().filter(|&&id| id == inst).count(), 1);
    }

    #[test]
    fn skipping_special_tokens_keeps_parser_markers() {
        let tokenizer = fixture();
        let id = |token: &str| id_of(&tokenizer, token).unwrap();
        let mut ids = vec![id("<s>"), id("[THINK]")];
        ids.extend(tokenizer.encode("plan").unwrap().token_ids());
        ids.extend([id("[/THINK]"), id("[TOOL_CALLS]")]);
        ids.extend(tokenizer.encode("f").unwrap().token_ids());
        ids.push(id("[ARGS]"));
        ids.extend(tokenizer.encode("{}").unwrap().token_ids());
        ids.push(id("</s>"));

        let skipped: String = tokenizer.decode(&ids, true).unwrap().into();
        assert_eq!(skipped, "[THINK]plan[/THINK][TOOL_CALLS]f[ARGS]{}");
        let kept: String = tokenizer.decode(&ids, false).unwrap().into();
        assert_eq!(kept, "<s>[THINK]plan[/THINK][TOOL_CALLS]f[ARGS]{}</s>");
    }

    #[test]
    fn decode_rejects_ids_outside_the_vocabulary() {
        let tokenizer = fixture();
        assert!(tokenizer.decode(&[1040], false).is_err());
    }

    #[test]
    fn decode_stream_reassembles_byte_fallback_characters() {
        let tokenizer = fixture();
        let ids = tokenizer.encode("北京 😀").unwrap().token_ids().to_vec();
        let tokenizer: Arc<dyn Tokenizer> = Arc::new(tokenizer);
        let mut stream = DecodeStream::new(tokenizer, &[], false);
        let mut text = String::new();
        for id in ids {
            if let Some(chunk) = stream.step(id).unwrap() {
                assert!(!chunk.contains('\u{FFFD}'), "chunk {chunk:?}");
                text.push_str(&chunk);
            }
        }
        if let Some(chunk) = stream.finish().unwrap() {
            text.push_str(&chunk);
        }
        assert_eq!(text, "北京 😀");
    }

    #[test]
    fn file_dispatch_selects_tekken() {
        let tokenizer = crate::create_tokenizer_from_file(FIXTURE).unwrap();
        assert_eq!(tokenizer.vocab_size(), Some(1040));
        let ids = tokenizer.encode("Hello, world!").unwrap();
        assert_eq!(
            ids.token_ids(),
            fixture().encode("Hello, world!").unwrap().token_ids()
        );
    }

    #[test]
    fn control_segments_match_special_token_matching() {
        let tokenizer = fixture();
        let fast = tokenizer
            .encode_segments(&[EncodeSegment::control("[INST]")])
            .unwrap();
        let mixed = tokenizer
            .encode_segments(&[EncodeSegment::control("[INST][/INST]")])
            .unwrap();
        assert_eq!(fast.token_ids(), [3]);
        assert_eq!(mixed.token_ids(), [3, 4]);
    }

    #[test]
    fn fallible_encoding_matches_core_bpe() {
        let tokenizer = fixture();
        for text in [
            "We're in Beijing 北京 😀\nDon't split café's accents.",
            "'re\u{0301} I'M 123\r\n\t  /[INST]/ [ARGS]{}",
            "<s>[INST]hi[/INST][TOOL_CALLS]f[ARGS]{}[THINK]plan[/THINK]",
        ] {
            assert_eq!(
                tokenizer.encode(text).unwrap().token_ids(),
                tokenizer.bpe.encode_ordinary(text)
            );
            assert_eq!(
                tokenizer
                    .encode_segments(&[EncodeSegment::control(text)])
                    .unwrap()
                    .token_ids(),
                tokenizer.bpe.encode_with_special_tokens(text)
            );
        }

        let long = "A".repeat(100_000);
        assert_eq!(
            tokenizer.encode(&long).unwrap().token_ids(),
            tokenizer.bpe.encode_ordinary(&long)
        );
    }

    #[test]
    fn long_regex_failures_are_returned_from_all_encoding_entry_points() {
        let tokenizer = fixture();
        for text in ["A".repeat(1_000_000), " ".repeat(1_000_000)] {
            let error = tokenizer.encode(&text).unwrap_err();
            assert!(error.to_string().contains("split regex"), "{error:#}");
            assert!(tokenizer.encode_batch(&[&text]).is_err());
            assert!(
                tokenizer
                    .encode_segments(&[EncodeSegment::ordinary(&text)])
                    .is_err()
            );
            let controlled = format!("<s>[INST]{text}[/INST]");
            assert!(
                tokenizer
                    .encode_segments(&[EncodeSegment::control(&controlled)])
                    .is_err()
            );
            assert!(tokenizer.token_to_id(&text).is_err());
        }
    }

    /// Load the fixture after `edit`, and return the load error.
    fn load_error(edit: impl FnOnce(&mut serde_json::Value)) -> String {
        let mut file: serde_json::Value =
            serde_json::from_str(&std::fs::read_to_string(FIXTURE).unwrap()).unwrap();
        edit(&mut file);
        let file: TekkenFile = serde_json::from_value(file).unwrap();
        TekkenTokenizer::from_parts(file).err().unwrap().to_string()
    }

    #[test]
    fn rejects_empty_special_tokens() {
        let error = load_error(|file| file["special_tokens"][5]["token_str"] = "".into());
        assert!(error.contains("empty"), "{error}");
    }

    #[test]
    fn rejects_vocab_without_the_byte_alphabet() {
        let error = load_error(|file| {
            let bytes = base64::engine::general_purpose::STANDARD.encode([0xff, 0xfe, 0xfd]);
            file["vocab"][65]["token_bytes"] = bytes.into();
        });
        assert!(error.contains("single byte 65"), "{error}");

        let error = load_error(|file| file["config"]["default_vocab_size"] = (40 + 255).into());
        assert!(error.contains("fewer than the 256"), "{error}");
    }

    #[test]
    fn rejects_oversized_special_token_reservation() {
        let error = load_error(|file| {
            file["config"]["default_vocab_size"] = u32::MAX.into();
            file["config"]["default_num_special_tokens"] = u32::MAX.into();
            file["special_tokens"] = serde_json::json!([]);
            file["vocab"] = serde_json::json!([]);
        });
        assert!(error.contains("special tokens, more than"), "{error}");
    }

    #[test]
    fn prefix_cache_is_refused() {
        assert!(fixture().validate_prefix_cache().is_err());
    }

    #[test]
    fn recognizes_tekken_file_names() {
        assert!(TekkenTokenizer::is_tekken_file(Path::new("/m/tekken.json")));
        assert!(TekkenTokenizer::is_tekken_file(Path::new(
            "tekken_240911.json"
        )));
        assert!(!TekkenTokenizer::is_tekken_file(Path::new(
            "/m/tokenizer.json"
        )));
        assert!(!TekkenTokenizer::is_tekken_file(Path::new(
            "/m/tekken.model"
        )));
    }
}