validation.rs 40.6 KB
Newer Older
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
1
/// Payload validation logic
OlivierDehaene's avatar
OlivierDehaene committed
2
use crate::config::Config;
3
use crate::validation::ValidationError::{BestOfSampling, BestOfSeed, EmptyInput};
4
5
use crate::{
    GenerateParameters, GenerateRequest, GrammarType, HubPreprocessorConfig, Idefics2Preprocessor,
6
    TokenizerTrait,
7
};
8
use crate::{PyTokenizer, Tokenizer};
OlivierDehaene's avatar
OlivierDehaene committed
9
use base64::{engine::general_purpose::STANDARD, Engine};
Nicolas Patry's avatar
Nicolas Patry committed
10
use image::{ImageFormat, ImageReader};
11
use jsonschema::{Draft, JSONSchema};
12
use outlines_core::json_schema::to_regex as json_schema_to_regex;
13
use rand::{thread_rng, Rng};
14
use serde_json::Value;
15
use std::io::Cursor;
16
use std::iter;
17
use std::sync::Arc;
Olivier Dehaene's avatar
Olivier Dehaene committed
18
use thiserror::Error;
OlivierDehaene's avatar
OlivierDehaene committed
19
use tokio::sync::mpsc;
20
use tokio::sync::oneshot;
21
use tracing::{instrument, Span};
22
use {once_cell::sync::Lazy, regex::Regex};
Olivier Dehaene's avatar
Olivier Dehaene committed
23

Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
24
/// Validation
Olivier Dehaene's avatar
Olivier Dehaene committed
25
#[derive(Debug, Clone)]
Olivier Dehaene's avatar
Olivier Dehaene committed
26
pub struct Validation {
27
    /// Validation parameters
28
    max_best_of: usize,
29
    max_stop_sequences: usize,
Nicolas Patry's avatar
Nicolas Patry committed
30
    max_top_n_tokens: u32,
31
32
    max_input_length: usize,
    max_total_tokens: usize,
drbh's avatar
drbh committed
33
    disable_grammar_support: bool,
34
    /// Channel to communicate with the background tokenization task
35
    sender: mpsc::UnboundedSender<TokenizerRequest>,
Olivier Dehaene's avatar
Olivier Dehaene committed
36
37
38
}

impl Validation {
OlivierDehaene's avatar
OlivierDehaene committed
39
    #[allow(clippy::too_many_arguments)]
40
41
    pub(crate) fn new(
        workers: usize,
42
        tokenizer: Tokenizer,
43
        config: Option<Config>,
44
        preprocessor_config: Option<HubPreprocessorConfig>,
45
        max_best_of: usize,
46
        max_stop_sequences: usize,
Nicolas Patry's avatar
Nicolas Patry committed
47
        max_top_n_tokens: u32,
48
49
        max_input_length: usize,
        max_total_tokens: usize,
drbh's avatar
drbh committed
50
        disable_grammar_support: bool,
51
    ) -> Self {
52
53
54
55
56
        let workers = if let Tokenizer::Python { .. } = &tokenizer {
            1
        } else {
            workers
        };
57
        // If we have a fast tokenizer
58
        let sender = {
OlivierDehaene's avatar
OlivierDehaene committed
59
60
61
            // Create round robin channel
            let (validation_sender, validation_round_robin_receiver) = mpsc::unbounded_channel();
            let mut senders = Vec::with_capacity(workers);
62
63
64
65

            // Create workers
            for _ in 0..workers {
                let tokenizer_clone = tokenizer.clone();
66
                let config_clone = config.clone();
67
                let preprocessor_config_clone = preprocessor_config.clone();
OlivierDehaene's avatar
OlivierDehaene committed
68
69
                let (tokenizer_sender, tokenizer_receiver) = mpsc::unbounded_channel();
                senders.push(tokenizer_sender);
70
71
72

                // Spawn worker
                tokio::task::spawn_blocking(move || {
73
74
75
76
77
78
                    tokenizer_worker(
                        tokenizer_clone,
                        config_clone,
                        preprocessor_config_clone,
                        tokenizer_receiver,
                    )
79
80
                });
            }
OlivierDehaene's avatar
OlivierDehaene committed
81
82
83
84

            // Create tokenization round robin task
            tokio::spawn(round_robin_task(validation_round_robin_receiver, senders));

85
            validation_sender
86
87
88
89
90
        };

        Self {
            max_best_of,
            sender,
91
            max_stop_sequences,
Nicolas Patry's avatar
Nicolas Patry committed
92
            max_top_n_tokens,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
93
            max_input_length,
94
            max_total_tokens,
drbh's avatar
drbh committed
95
            disable_grammar_support,
96
97
        }
    }
Olivier Dehaene's avatar
Olivier Dehaene committed
98

99
    #[instrument(skip(self, inputs))]
100
    pub async fn tokenize(
101
102
        &self,
        inputs: String,
103
        add_special_tokens: bool,
104
        truncate: Option<usize>,
105
    ) -> Result<(tokenizers::Encoding, Vec<Chunk>), ValidationError> {
106
        // If we have a fast tokenizer
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
        // Create response channel
        let (response_sender, response_receiver) = oneshot::channel();
        // Send request to the background validation task
        // Unwrap is safe here
        let _ = &self
            .sender
            .send((
                (inputs, add_special_tokens, truncate),
                response_sender,
                Span::current(),
            ))
            .unwrap();

        // Await on response channel
        // Unwrap is safe here
        let encoding = response_receiver.await.unwrap()?;
        Ok(encoding)
124
125
    }

126
    #[allow(clippy::type_complexity)]
127
128
129
130
    #[instrument(skip(self, inputs))]
    async fn validate_input(
        &self,
        inputs: String,
131
        add_special_tokens: bool,
132
133
        truncate: Option<usize>,
        max_new_tokens: Option<u32>,
134
    ) -> Result<(Vec<Chunk>, Option<Vec<u32>>, usize, u32), ValidationError> {
135
        // If we have a fast tokenizer
136
        let (encoding, inputs) = self
137
            .tokenize(inputs.clone(), add_special_tokens, truncate)
138
139
140
141
142
143
144
            .await?;
        // Create response channel
        let input_length = if let Some(truncate) = truncate {
            std::cmp::min(encoding.len(), truncate)
        } else {
            encoding.len()
        };
145

146
147
148
149
150
151
152
        // Get total tokens
        let max_new_tokens: u32 = if let Some(max_new_tokens) = max_new_tokens {
            max_new_tokens
        } else {
            self.max_total_tokens.saturating_sub(input_length) as u32
        };
        let total_tokens = input_length + max_new_tokens as usize;
153

154
155
156
157
158
159
160
        // Validate MaxTotalTokens
        if total_tokens > self.max_total_tokens {
            return Err(ValidationError::MaxTotalTokens(
                self.max_total_tokens,
                input_length,
                max_new_tokens,
            ));
161
162
        }

163
164
165
166
        // Validate InputLength
        if input_length > self.max_input_length {
            return Err(ValidationError::InputLength(
                self.max_input_length,
167
                input_length,
168
            ));
Olivier Dehaene's avatar
Olivier Dehaene committed
169
        }
170
171
172
173
174
175

        let ids = encoding.get_ids();
        let input_ids = ids[ids.len().saturating_sub(input_length)..].to_owned();

        metrics::histogram!("tgi_request_input_length").record(input_length as f64);
        Ok((inputs, Some(input_ids), input_length, max_new_tokens))
Olivier Dehaene's avatar
Olivier Dehaene committed
176
177
    }

Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
178
    /// Validate a payload and get the number of tokens in the input
179
    #[instrument(skip_all)]
Olivier Dehaene's avatar
Olivier Dehaene committed
180
181
182
    pub(crate) async fn validate(
        &self,
        request: GenerateRequest,
183
    ) -> Result<ValidGenerateRequest, ValidationError> {
184
185
186
187
        let GenerateParameters {
            best_of,
            temperature,
            repetition_penalty,
188
            frequency_penalty,
189
190
191
192
193
194
195
196
197
            top_k,
            top_p,
            typical_p,
            do_sample,
            max_new_tokens,
            stop: stop_sequences,
            truncate,
            seed,
            watermark,
198
            decoder_input_details,
Nicolas Patry's avatar
Nicolas Patry committed
199
            top_n_tokens,
drbh's avatar
drbh committed
200
            grammar,
drbh's avatar
drbh committed
201
            adapter_id,
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
            ..
        } = request.parameters;

        // sampling must be true when best_of > 1
        let best_of = best_of.unwrap_or(1);
        let sampling = do_sample
            || temperature.is_some()
            || top_k.is_some()
            || top_p.is_some()
            || typical_p.is_some();

        if best_of > 1 && !sampling {
            return Err(BestOfSampling);
        }

        let temperature = temperature.unwrap_or(1.0);
        if temperature <= 0.0 {
            return Err(ValidationError::Temperature);
        }

        let repetition_penalty = repetition_penalty.unwrap_or(1.0);
        if repetition_penalty <= 0.0 {
            return Err(ValidationError::RepetitionPenalty);
        }

227
228
229
230
231
        let frequency_penalty = frequency_penalty.unwrap_or(0.0);
        if !(-2.0..=2.0).contains(&frequency_penalty) {
            return Err(ValidationError::FrequencyPenalty);
        }

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
        // Different because the proto default value is not a valid value
        // for the user
        let top_p = top_p
            .map(|value| {
                if value <= 0.0 || value >= 1.0 {
                    return Err(ValidationError::TopP);
                }
                Ok(value)
            })
            .unwrap_or(Ok(1.0))?;

        let typical_p = typical_p
            .map(|value| {
                if value <= 0.0 || value >= 1.0 {
                    return Err(ValidationError::TypicalP);
                }
                Ok(value)
            })
            .unwrap_or(Ok(1.0))?;

        let top_k: u32 = top_k
            .map(|value| {
                if value <= 0 {
                    return Err(ValidationError::TopK);
                }
                Ok(value as u32)
            })
            .unwrap_or(Ok(0))?;

261
        if max_new_tokens == Some(0) {
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
            return Err(ValidationError::NegativeMaxNewTokens);
        }

        if stop_sequences.len() > self.max_stop_sequences {
            return Err(ValidationError::StopSequence(
                self.max_stop_sequences,
                stop_sequences.len(),
            ));
        }

        // If seed is None, assign a random one
        let seed = match seed {
            None => thread_rng().gen(),
            Some(seed) => {
                if best_of > 1 {
                    return Err(BestOfSeed);
                }
                seed
            }
        };

Nicolas Patry's avatar
Nicolas Patry committed
283
284
285
286
287
288
289
290
291
        let top_n_tokens = top_n_tokens
            .map(|value| {
                if value > self.max_top_n_tokens {
                    return Err(ValidationError::TopNTokens(self.max_top_n_tokens, value));
                }
                Ok(value)
            })
            .unwrap_or(Ok(0))?;

292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
        // Check if inputs is empty
        if request.inputs.is_empty() {
            return Err(EmptyInput);
        }

        // Check if truncate is strictly positive and less than max_input_length
        let truncate = truncate
            .map(|value| {
                if value == 0 || value > self.max_input_length {
                    return Err(ValidationError::Truncate(self.max_input_length, value));
                }
                Ok(Some(value))
            })
            .unwrap_or(Ok(None))?;

        // Validate inputs
308
        let (inputs, input_ids, input_length, max_new_tokens) = self
309
310
311
312
313
314
            .validate_input(
                request.inputs,
                request.add_special_tokens,
                truncate,
                max_new_tokens,
            )
315
316
            .await?;

drbh's avatar
drbh committed
317
318
319
320
321
322
323
        // TODO: we should build the FSM here and pass the compiled FSM instead of the grammar
        // NOTE: this is currently difficult because we need the tokenizer in Python to build
        // the FSM and we'd have to load a copy of the tokenizer into our Pyo3 instance which
        // may be slow and memory intensive. Best case is to have a Rust implementation of the FSM
        // compiler and use that to build the FSM here.

        // Validate grammar and unpack the grammar and type for the proto message
OlivierDehaene's avatar
OlivierDehaene committed
324
        let grammar = match grammar {
drbh's avatar
drbh committed
325
326
327
328
329
            Some(grammar) => {
                // Ensure that grammar is not set if it's not supported
                if self.disable_grammar_support {
                    return Err(ValidationError::Grammar);
                }
OlivierDehaene's avatar
OlivierDehaene committed
330
                let valid_grammar = match grammar {
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
                    GrammarType::Json(json) => {
                        let json = match json {
                            // if value is a string, we need to parse it again to make sure its
                            // a valid json
                            Value::String(s) => serde_json::from_str(&s)
                                .map_err(|e| ValidationError::InvalidGrammar(e.to_string())),
                            Value::Object(_) => Ok(json),
                            _ => Err(ValidationError::Grammar),
                        }?;

                        // Check if the json is a valid JSONSchema
                        JSONSchema::options()
                            .with_draft(Draft::Draft202012)
                            .compile(&json)
                            .map_err(|e| ValidationError::InvalidGrammar(e.to_string()))?;

347
348
349
350
351
352
353
354
                        // The schema can be valid but lack properties.
                        // We need properties for the grammar to be successfully parsed in Python.
                        // Therefore, we must check and throw an error if properties are missing.
                        json.get("properties")
                            .ok_or(ValidationError::InvalidGrammar(
                                "Grammar must have a 'properties' field".to_string(),
                            ))?;

355
356
357
358
359
360
361
                        // Do compilation in the router for performance. In the future, we
                        // should also move regex -> automaton compilation in the router,
                        // but this is not yet supported in pure Rust by outlines-core.
                        let grammar_regex = json_schema_to_regex(&json, None, &json)
                            .map_err(ValidationError::RegexFromSchema)?;

                        ValidGrammar::Regex(grammar_regex.to_string())
362
                    }
OlivierDehaene's avatar
OlivierDehaene committed
363
364
365
                    GrammarType::Regex(regex) => ValidGrammar::Regex(regex),
                };
                Some(valid_grammar)
drbh's avatar
drbh committed
366
            }
OlivierDehaene's avatar
OlivierDehaene committed
367
            None => None,
drbh's avatar
drbh committed
368
369
        };

OlivierDehaene's avatar
OlivierDehaene committed
370
        let parameters = ValidParameters {
371
372
            temperature,
            repetition_penalty,
373
            frequency_penalty,
374
375
376
377
378
379
            top_k,
            top_p,
            typical_p,
            do_sample,
            seed,
            watermark,
drbh's avatar
drbh committed
380
            grammar,
381
        };
OlivierDehaene's avatar
OlivierDehaene committed
382
        let stopping_parameters = ValidStoppingParameters {
383
384
385
386
387
            max_new_tokens,
            stop_sequences,
            ignore_eos_token: false,
        };

388
        metrics::histogram!("tgi_request_max_new_tokens").record(max_new_tokens as f64);
389
390
391

        Ok(ValidGenerateRequest {
            inputs,
392
            input_ids: input_ids.map(Arc::new),
393
            add_special_tokens: request.add_special_tokens,
394
            decoder_input_details,
395
            input_length: input_length as u32,
396
397
398
            truncate: truncate.unwrap_or(self.max_input_length) as u32,
            parameters,
            stopping_parameters,
Nicolas Patry's avatar
Nicolas Patry committed
399
            top_n_tokens,
drbh's avatar
drbh committed
400
            adapter_id,
401
        })
Olivier Dehaene's avatar
Olivier Dehaene committed
402
    }
403
404
405
406
407
408
409
410
411
412
413
414
415
416

    /// Validate the best_of parameter
    #[instrument(skip_all)]
    pub(crate) fn validate_best_of(&self, best_of: usize) -> Result<usize, ValidationError> {
        if self.max_best_of == 1 && best_of != 1 {
            return Err(ValidationError::BestOfDisabled);
        }

        if best_of > self.max_best_of {
            return Err(ValidationError::BestOf(self.max_best_of, best_of));
        }

        Ok(best_of)
    }
Olivier Dehaene's avatar
Olivier Dehaene committed
417
418
}

OlivierDehaene's avatar
OlivierDehaene committed
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
/// Round robin tokenization task
async fn round_robin_task(
    mut receiver: mpsc::UnboundedReceiver<TokenizerRequest>,
    senders: Vec<mpsc::UnboundedSender<TokenizerRequest>>,
) {
    loop {
        for sender in &senders {
            match receiver.recv().await {
                None => return,
                Some(request) => sender.send(request).unwrap(),
            };
        }
    }
}

434
/// Start tokenization workers
435
436
437
fn tokenizer_worker(
    tokenizer: Tokenizer,
    config: Option<Config>,
438
    preprocessor_config: Option<HubPreprocessorConfig>,
439
440
    mut receiver: mpsc::UnboundedReceiver<TokenizerRequest>,
) {
441
442
443
444
    match tokenizer {
        Tokenizer::Python {
            tokenizer_name,
            revision,
445
            trust_remote_code,
446
447
        } => {
            pyo3::Python::with_gil(|py| -> pyo3::PyResult<()> {
448
449
                let tokenizer =
                    PyTokenizer::from_py(py, tokenizer_name, revision, trust_remote_code)?;
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
                // Loop over requests
                while let Some(((inputs, add_special_tokens, truncate), response_tx, parent_span)) =
                    receiver.blocking_recv()
                {
                    parent_span.in_scope(|| {
                        response_tx
                            .send(prepare_input(
                                inputs,
                                truncate,
                                add_special_tokens,
                                &tokenizer,
                                config.as_ref(),
                                preprocessor_config.as_ref(),
                            ))
                            .unwrap_or(())
                    })
                }
                Ok(())
            })
            .expect("Failure in python tokenizer worker");
        }
        Tokenizer::Rust(tokenizer) => {
            while let Some(((inputs, add_special_tokens, truncate), response_tx, parent_span)) =
                receiver.blocking_recv()
            {
                parent_span.in_scope(|| {
                    response_tx
                        .send(prepare_input(
                            inputs,
                            truncate,
                            add_special_tokens,
                            &tokenizer,
                            config.as_ref(),
                            preprocessor_config.as_ref(),
                        ))
                        .unwrap_or(())
                })
            }
        }
489
490
    }
}
Olivier Dehaene's avatar
Olivier Dehaene committed
491

492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
fn format_from_mimetype(mimetype: &str) -> Option<ImageFormat> {
    match mimetype {
        "image/png" => Some(ImageFormat::Png),
        "image/jpeg" => Some(ImageFormat::Jpeg),
        "image/jpg" => Some(ImageFormat::Jpeg),
        "image/gif" => Some(ImageFormat::Gif),
        "image/webp" => Some(ImageFormat::WebP),
        "image/tiff" => Some(ImageFormat::Tiff),
        // "image/pnm"=>Some(ImageFormat::Pnm),
        // "image/tga"=>Some(ImageFormat::Tga),
        // "image/dds"=>Some(ImageFormat::Dds),
        // "image/bmp"=>Some(ImageFormat::Bmp),
        // "image/ico"=>Some(ImageFormat::Ico),
        // "image/x-exr"=>Some(ImageFormat::OpenExr),
        _ => None,
    }
}
OlivierDehaene's avatar
OlivierDehaene committed
509

510
511
512
513
514
515
516
517
518
519
520
521
fn format_to_mimetype(format: ImageFormat) -> String {
    match format {
        ImageFormat::Png => "image/png",
        ImageFormat::Jpeg => "image/jpeg",
        ImageFormat::Gif => "image/gif",
        ImageFormat::WebP => "image/webp",
        ImageFormat::Tiff => "image/tiff",
        _ => "application/octet-stream",
    }
    .to_string()
}

522
fn fetch_image(input: &str) -> Result<(Vec<u8>, String, usize, usize), ValidationError> {
523
524
525
526
527
528
529
530
531
532
    if input.starts_with("![](http://") || input.starts_with("![](https://") {
        let url = &input["![](".len()..input.len() - 1];
        let data = reqwest::blocking::get(url)?.bytes()?;

        let format = image::guess_format(&data)?;
        // TODO Remove this clone
        let img = ImageReader::with_format(Cursor::new(data.clone()), format).decode()?;
        let height: usize = img.height().try_into()?;
        let width: usize = img.width().try_into()?;
        let mimetype = format_to_mimetype(format);
533
        Ok((data.to_vec(), mimetype, height, width))
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
    } else if input.starts_with("![](data:") {
        // Remove ![](....)
        let content = &input["![](data:".len()..input.len() - 1];
        let tokens: Vec<_> = content.split(';').collect();
        if tokens.len() != 2 {
            return Err(ValidationError::InvalidImageContent(content.to_string()));
        }
        let mimetype = tokens[0];
        let content = tokens[1];

        if !content.starts_with("base64,") {
            return Err(ValidationError::InvalidImageContent(content.to_string()));
        }

        let data = STANDARD.decode(content["base64,".len()..].as_bytes())?;
        let img = if let Some(format) = format_from_mimetype(mimetype) {
550
            ImageReader::with_format(Cursor::new(&data), format).decode()?
551
        } else {
552
            ImageReader::new(Cursor::new(&data))
553
554
555
556
557
558
559
                .with_guessed_format()
                .map_err(|_io_error| ValidationError::InvalidImageContent(content.to_string()))?
                .decode()?
        };

        let height: usize = img.height().try_into()?;
        let width: usize = img.width().try_into()?;
560
        Ok((data, mimetype.to_string(), height, width))
561
562
563
564
565
    } else {
        Err(ValidationError::InvalidImageContent(input.to_string()))
    }
}

566
567
568
569
570
571
572
573
574
575
fn image_tokens(
    config: &Config,
    preprocessor_config: Option<&HubPreprocessorConfig>,
    height: usize,
    width: usize,
) -> String {
    use Config::*;
    use HubPreprocessorConfig::*;
    match config {
        Idefics => "<image>".to_string(),
Nicolas Patry's avatar
Nicolas Patry committed
576
        Mllama => "<|image|>".to_string(),
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
        Idefics2(config) => {
            const FAKE: &str = "<fake_token_around_image>";
            const IMAGE: &str = "<image>";

            let slots = config.get_number_of_features(height, width);

            let mut image_string = String::with_capacity(2 * FAKE.len() + slots * IMAGE.len());
            image_string.push_str(FAKE);
            image_string.extend(iter::repeat(IMAGE).take(slots));
            image_string.push_str(FAKE);

            if matches!(
                preprocessor_config,
                Some(Idefics2Processor(Idefics2Preprocessor {
                    do_image_splitting: true,
                    ..
                }))
            ) {
                image_string = image_string.repeat(5);
            };

            image_string
        }
        Paligemma(config) => "<image>".repeat(config.get_number_of_features(height, width)),
        LlavaNext(config) => "<image>".repeat(config.get_number_of_features(height, width)),
drbh's avatar
drbh committed
602
603
604
605
        Qwen2Vl(config) => format!(
            "<|vision_start|>{:?}<|vision_end|>",
            "<|image_pad|>".repeat(config.get_number_of_features(height, width))
        ),
606
607
608
609
610
611
612
613
614
615
616
617
618
619
        _ => unimplemented!("Images tokens are not supported for this model configuration"),
    }
}

fn image_tokens_fixup(config: &Config, text: String) -> String {
    match config {
        Config::Idefics2(_) => {
            const FAKE: &str = "<fake_token_around_image>";
            text.replace(&format!("{FAKE}{FAKE}"), FAKE)
        }
        _ => text,
    }
}

620
/// Get input length and optionally truncate it
621
fn prepare_input<T: TokenizerTrait>(
622
    inputs: String,
623
    _truncate: Option<usize>,
624
    add_special_tokens: bool,
625
    tokenizer: &T,
626
627
    config: Option<&Config>,
    preprocessor_config: Option<&HubPreprocessorConfig>,
Nicolas Patry's avatar
Nicolas Patry committed
628
) -> Result<(tokenizers::Encoding, Vec<Chunk>), ValidationError> {
629
    use Config::*;
630
    static RE: Lazy<Regex> = Lazy::new(|| Regex::new(r"!\[\]\([^\)]*\)").unwrap());
631
    let (tokenizer_query, input_chunks) = match config {
drbh's avatar
drbh committed
632
633
634
        Some(
            config @ (Idefics | Mllama | Idefics2(_) | Paligemma(_) | LlavaNext(_) | Qwen2Vl(_)),
        ) => {
635
            let mut input_chunks = Vec::new();
636
637
638
639
            let mut tokenizer_query = String::with_capacity(inputs.len());
            let mut start = 0;
            for chunk in RE.find_iter(&inputs) {
                let chunk_start = chunk.start();
drbh's avatar
drbh committed
640
641
                let chunk_end = chunk.end();
                if chunk_start != start {
Nicolas Patry's avatar
Nicolas Patry committed
642
                    input_chunks.push(Chunk::Text(inputs[start..chunk_start].to_string()));
drbh's avatar
drbh committed
643
644
                    tokenizer_query.push_str(&inputs[start..chunk_start]);
                }
645
                let (data, mimetype, height, width) = fetch_image(&inputs[chunk_start..chunk_end])?;
Nicolas Patry's avatar
Nicolas Patry committed
646
                input_chunks.push(Chunk::Image(Image { data, mimetype }));
647
                tokenizer_query.push_str(&image_tokens(config, preprocessor_config, height, width));
drbh's avatar
drbh committed
648
649
                start = chunk_end;
            }
650
            if start != inputs.len() {
Nicolas Patry's avatar
Nicolas Patry committed
651
                input_chunks.push(Chunk::Text(inputs[start..].to_string()));
drbh's avatar
drbh committed
652
653
                tokenizer_query.push_str(&inputs[start..]);
            }
Nicolas Patry's avatar
Nicolas Patry committed
654

655
656
            tokenizer_query = image_tokens_fixup(config, tokenizer_query);

657
            (tokenizer_query, input_chunks)
Nicolas Patry's avatar
Nicolas Patry committed
658
        }
Nicolas Patry's avatar
Nicolas Patry committed
659
        _ => (inputs.clone(), vec![Chunk::Text(inputs)]),
660
    };
661

662
    // Get the number of tokens in the input
663
    let encoding = tokenizer
664
        .encode_trait(tokenizer_query, add_special_tokens)
665
666
        .map_err(|err| ValidationError::Tokenizer(err.to_string()))?;

667
    Ok((encoding, input_chunks))
Olivier Dehaene's avatar
Olivier Dehaene committed
668
}
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
669

670
type TokenizerRequest = (
671
    (String, bool, Option<usize>),
Nicolas Patry's avatar
Nicolas Patry committed
672
    oneshot::Sender<Result<(tokenizers::Encoding, Vec<Chunk>), ValidationError>>,
673
    Span,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
674
675
);

Nicolas Patry's avatar
Nicolas Patry committed
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
#[derive(Debug, Clone, Eq, PartialEq)]
pub struct Image {
    pub data: Vec<u8>,
    pub mimetype: String,
}

#[derive(Debug, Clone, Eq, PartialEq)]
pub enum Chunk {
    Text(String),
    Image(Image),
}

/// Convert input chunks to a stringly-typed input for backwards
/// compat for backends that haven't implemented chunked inputs.
pub trait ChunksToString {
    /// Convert chunks to string.
    fn chunks_to_string(&self) -> String;
}

impl ChunksToString for Vec<Chunk> {
    fn chunks_to_string(&self) -> String {
        let mut output = String::new();
        self.iter().for_each(|c| match &c {
            Chunk::Text(text) => output.push_str(text),
            Chunk::Image(Image { data, mimetype }) => {
                let encoded = STANDARD.encode(data);
                output.push_str(&format!("![](data:{};base64,{})", mimetype, encoded))
            }
        });
        output
    }
}

OlivierDehaene's avatar
OlivierDehaene committed
709
#[derive(Debug, Clone)]
Nicolas Patry's avatar
Nicolas Patry committed
710
pub enum ValidGrammar {
OlivierDehaene's avatar
OlivierDehaene committed
711
712
713
714
715
    Json(String),
    Regex(String),
}

#[derive(Debug, Clone)]
Nicolas Patry's avatar
Nicolas Patry committed
716
pub struct ValidParameters {
OlivierDehaene's avatar
OlivierDehaene committed
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
    /// / exponential scaling output probability distribution
    pub temperature: f32,
    /// / restricting to the k highest probability elements
    pub top_k: u32,
    /// / restricting to top tokens summing to prob_cut_off <= prob_cut_off
    pub top_p: f32,
    /// / restricting to top tokens summing to prob_cut_off <= prob_cut_off
    pub typical_p: f32,
    /// / apply sampling on the logits
    pub do_sample: bool,
    /// / random seed for sampling
    pub seed: u64,
    /// / repetition penalty
    pub repetition_penalty: f32,
    /// / frequency penalty
    pub frequency_penalty: f32,
    /// / token watermarking using "A Watermark for Large Language Models"
    pub watermark: bool,
    /// / grammar (applied if not empty)
    pub grammar: Option<ValidGrammar>,
}

#[derive(Debug, Clone)]
Nicolas Patry's avatar
Nicolas Patry committed
740
pub struct ValidStoppingParameters {
OlivierDehaene's avatar
OlivierDehaene committed
741
742
743
744
745
746
747
748
749
    /// / Maximum number of generated tokens
    pub max_new_tokens: u32,
    /// / Optional stopping sequences
    pub stop_sequences: Vec<String>,
    /// / Ignore end of sequence token
    /// / used for benchmarking
    pub ignore_eos_token: bool,
}

750
#[derive(Debug, Clone)]
Nicolas Patry's avatar
Nicolas Patry committed
751
752
pub struct ValidGenerateRequest {
    pub inputs: Vec<Chunk>,
753
    pub input_ids: Option<Arc<Vec<u32>>>,
754
    pub input_length: u32,
755
    pub truncate: u32,
756
    pub add_special_tokens: bool,
757
    pub decoder_input_details: bool,
OlivierDehaene's avatar
OlivierDehaene committed
758
759
    pub parameters: ValidParameters,
    pub stopping_parameters: ValidStoppingParameters,
Nicolas Patry's avatar
Nicolas Patry committed
760
    pub top_n_tokens: u32,
drbh's avatar
drbh committed
761
    pub adapter_id: Option<String>,
762
763
}

Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
764
765
#[derive(Error, Debug)]
pub enum ValidationError {
766
767
768
769
770
771
772
773
774
775
    #[error("`best_of` must be > 0 and <= {0}. Given: {1}")]
    BestOf(usize, usize),
    #[error("`best_of` != 1 is not allowed for this endpoint")]
    BestOfDisabled,
    #[error("you must use sampling when `best_of` is > 1")]
    BestOfSampling,
    #[error("`seed` must not be set when `best_of` > 1")]
    BestOfSeed,
    #[error("`best_of` != 1 is not supported when streaming tokens")]
    BestOfStream,
Nicolas Patry's avatar
Nicolas Patry committed
776
777
778
779
    #[error("`top_n_tokens` must be >= 0 and <= {0}. Given: {1}")]
    TopNTokens(u32, u32),
    #[error("`top_n_tokens` != 0 is not allowed for this endpoint")]
    TopNTokensDisabled,
780
781
    #[error("`decoder_input_details` == true is not supported when streaming tokens")]
    PrefillDetailsStream,
782
    #[error("`temperature` must be strictly positive")]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
783
    Temperature,
784
    #[error("`repetition_penalty` must be strictly positive")]
785
    RepetitionPenalty,
786
787
    #[error("`frequency_penalty` must be >= -2.0 and <= 2.0")]
    FrequencyPenalty,
788
    #[error("`top_p` must be > 0.0 and < 1.0")]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
789
    TopP,
790
    #[error("`top_k` must be strictly positive")]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
791
    TopK,
792
793
    #[error("`truncate` must be strictly positive and less than {0}. Given: {1}")]
    Truncate(usize, usize),
794
795
    #[error("`typical_p` must be > 0.0 and < 1.0")]
    TypicalP,
796
797
    #[error("one of `max_new_tokens` or `truncate` must be set if a fast tokenizer is not in use")]
    UnsetMaxNewTokens,
798
    #[error("`max_new_tokens` must be strictly positive")]
799
800
801
    NegativeMaxNewTokens,
    #[error("`max_new_tokens` must be <= {0}. Given: {1}")]
    MaxNewTokens(usize, u32),
802
    #[error("`inputs` tokens + `max_new_tokens` must be <= {0}. Given: {1} `inputs` tokens and {2} `max_new_tokens`")]
803
    MaxTotalTokens(usize, usize, u32),
804
    #[error("`inputs` must have less than {0} tokens. Given: {1}")]
805
    InputLength(usize, usize),
806
    #[error("`inputs` cannot be empty")]
807
    EmptyInput,
808
    #[error("`stop` supports up to {0} stop sequences. Given: {1}")]
809
    StopSequence(usize, usize),
810
811
    #[error("tokenizer error {0}")]
    Tokenizer(String),
drbh's avatar
drbh committed
812
813
    #[error("grammar is not supported")]
    Grammar,
814
815
    #[error("grammar is not valid: {0}")]
    InvalidGrammar(String),
816
817
    #[error("cannot compile regex from schema: {0}")]
    RegexFromSchema(anyhow::Error),
818
819
820
821
822
823
824
825
826
827
    #[error("base64 encoding is invalid: {0}")]
    InvalidBase64(#[from] base64::DecodeError),
    #[error("invalid image: {0}")]
    InvalidImage(#[from] image::ImageError),
    #[error("invalid integer: {0}")]
    InvalidInt(#[from] core::num::TryFromIntError),
    #[error("invalid image content: {0}")]
    InvalidImageContent(String),
    #[error("Could not fetch image: {0}")]
    FailedFetchImage(#[from] reqwest::Error),
Nicolas Patry's avatar
Nicolas Patry committed
828
829
    #[error("{0} modality is not supported")]
    UnsupportedModality(&'static str),
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
830
}
831
832

#[cfg(test)]
833
mod tests {
834
    use super::*;
835
    use crate::config::{Idefics2, PaliTextConfig, Paligemma};
836
837
    use crate::default_parameters;
    use crate::tests::get_tokenizer;
838
839

    #[tokio::test]
840
    async fn test_validation_max_new_tokens() {
841
        let tokenizer = get_tokenizer();
842
843
        let max_best_of = 2;
        let max_stop_sequence = 3;
Nicolas Patry's avatar
Nicolas Patry committed
844
845
846
        let max_top_n_tokens = 4;
        let max_input_length = 5;
        let max_total_tokens = 6;
847
        let workers = 1;
drbh's avatar
drbh committed
848
        let disable_grammar_support = true;
849
        let config = None;
850
851
852
        let validation = Validation::new(
            workers,
            tokenizer,
853
            config,
854
            None,
855
856
            max_best_of,
            max_stop_sequence,
Nicolas Patry's avatar
Nicolas Patry committed
857
            max_top_n_tokens,
858
859
            max_input_length,
            max_total_tokens,
drbh's avatar
drbh committed
860
            disable_grammar_support,
861
        );
862
863

        let max_new_tokens = 10;
864
        match validation
865
            .validate_input("Hello".to_string(), true, None, Some(max_new_tokens))
866
867
            .await
        {
868
869
            Err(ValidationError::MaxTotalTokens(6, 1, 10)) => (),
            // Ok((_s, _, 0, 10)) => (),
870
            r => panic!("Unexpected not max new tokens: {r:?}"),
871
872
873
874
        }
    }

    #[tokio::test]
875
    async fn test_validation_input_length() {
876
        let tokenizer = get_tokenizer();
877
878
        let max_best_of = 2;
        let max_stop_sequence = 3;
Nicolas Patry's avatar
Nicolas Patry committed
879
880
881
        let max_top_n_tokens = 4;
        let max_input_length = 5;
        let max_total_tokens = 6;
drbh's avatar
drbh committed
882
        let disable_grammar_support = true;
883
        let workers = 1;
884
        let config = None;
885
886
887
        let validation = Validation::new(
            workers,
            tokenizer,
888
            config,
889
            None,
890
891
            max_best_of,
            max_stop_sequence,
Nicolas Patry's avatar
Nicolas Patry committed
892
            max_top_n_tokens,
893
894
            max_input_length,
            max_total_tokens,
drbh's avatar
drbh committed
895
            disable_grammar_support,
896
        );
897
898

        let max_new_tokens = 10;
899
        match validation
900
            .validate_input("Hello".to_string(), true, None, Some(max_new_tokens))
901
902
            .await
        {
Nicolas Patry's avatar
Nicolas Patry committed
903
            Err(ValidationError::MaxTotalTokens(6, 1, 10)) => (),
904
            _ => panic!("Unexpected not max new tokens"),
905
906
        }
    }
907
908

    #[tokio::test]
909
    async fn test_validation_best_of_sampling() {
910
        let tokenizer = get_tokenizer();
911
912
        let max_best_of = 2;
        let max_stop_sequence = 3;
Nicolas Patry's avatar
Nicolas Patry committed
913
914
915
        let max_top_n_tokens = 4;
        let max_input_length = 5;
        let max_total_tokens = 6;
916
        let workers = 1;
drbh's avatar
drbh committed
917
        let disable_grammar_support = true;
918
        let config = None;
919
920
921
        let validation = Validation::new(
            workers,
            tokenizer,
922
            config,
923
            None,
924
925
            max_best_of,
            max_stop_sequence,
Nicolas Patry's avatar
Nicolas Patry committed
926
            max_top_n_tokens,
927
928
            max_input_length,
            max_total_tokens,
drbh's avatar
drbh committed
929
            disable_grammar_support,
930
931
932
933
        );
        match validation
            .validate(GenerateRequest {
                inputs: "Hello".to_string(),
934
                add_special_tokens: true,
935
936
937
938
939
940
941
942
                parameters: GenerateParameters {
                    best_of: Some(2),
                    do_sample: false,
                    ..default_parameters()
                },
            })
            .await
        {
943
            Err(ValidationError::BestOfSampling) => (),
944
            _ => panic!("Unexpected not best of sampling"),
945
946
947
948
        }
    }

    #[tokio::test]
949
    async fn test_validation_top_p() {
950
        let tokenizer = get_tokenizer();
951
952
        let max_best_of = 2;
        let max_stop_sequence = 3;
Nicolas Patry's avatar
Nicolas Patry committed
953
954
        let max_top_n_tokens = 4;
        let max_input_length = 5;
955
        let max_total_tokens = 106;
956
        let workers = 1;
drbh's avatar
drbh committed
957
        let disable_grammar_support = true;
958
        let config = None;
959
960
961
        let validation = Validation::new(
            workers,
            tokenizer,
962
            config,
963
            None,
964
965
            max_best_of,
            max_stop_sequence,
Nicolas Patry's avatar
Nicolas Patry committed
966
            max_top_n_tokens,
967
968
            max_input_length,
            max_total_tokens,
drbh's avatar
drbh committed
969
            disable_grammar_support,
970
971
972
973
        );
        match validation
            .validate(GenerateRequest {
                inputs: "Hello".to_string(),
974
                add_special_tokens: true,
975
976
                parameters: GenerateParameters {
                    top_p: Some(1.0),
977
                    max_new_tokens: Some(5),
978
979
980
981
982
                    ..default_parameters()
                },
            })
            .await
        {
983
            Err(ValidationError::TopP) => (),
984
            _ => panic!("Unexpected top_p"),
985
986
        }

987
988
989
        match validation
            .validate(GenerateRequest {
                inputs: "Hello".to_string(),
990
                add_special_tokens: true,
991
992
                parameters: GenerateParameters {
                    top_p: Some(0.99),
993
                    max_new_tokens: Some(5),
994
995
996
997
998
                    ..default_parameters()
                },
            })
            .await
        {
999
            Ok(_) => (),
1000
            _ => panic!("Unexpected top_p error"),
1001
1002
        }

1003
1004
1005
        let valid_request = validation
            .validate(GenerateRequest {
                inputs: "Hello".to_string(),
1006
                add_special_tokens: true,
1007
1008
                parameters: GenerateParameters {
                    top_p: None,
1009
                    max_new_tokens: Some(5),
1010
1011
1012
1013
1014
                    ..default_parameters()
                },
            })
            .await
            .unwrap();
1015
1016
1017
        // top_p == 1.0 is invalid for users to ask for but it's the default resolved value.
        assert_eq!(valid_request.parameters.top_p, 1.0);
    }
Nicolas Patry's avatar
Nicolas Patry committed
1018
1019
1020

    #[tokio::test]
    async fn test_validation_top_n_tokens() {
1021
        let tokenizer = get_tokenizer();
Nicolas Patry's avatar
Nicolas Patry committed
1022
1023
1024
1025
        let max_best_of = 2;
        let max_stop_sequences = 3;
        let max_top_n_tokens = 4;
        let max_input_length = 5;
1026
        let max_total_tokens = 106;
Nicolas Patry's avatar
Nicolas Patry committed
1027
        let workers = 1;
drbh's avatar
drbh committed
1028
        let disable_grammar_support = true;
1029
        let config = None;
Nicolas Patry's avatar
Nicolas Patry committed
1030
1031
1032
        let validation = Validation::new(
            workers,
            tokenizer,
1033
            config,
1034
            None,
Nicolas Patry's avatar
Nicolas Patry committed
1035
1036
1037
1038
1039
            max_best_of,
            max_stop_sequences,
            max_top_n_tokens,
            max_input_length,
            max_total_tokens,
drbh's avatar
drbh committed
1040
            disable_grammar_support,
Nicolas Patry's avatar
Nicolas Patry committed
1041
1042
1043
1044
        );
        match validation
            .validate(GenerateRequest {
                inputs: "Hello".to_string(),
1045
                add_special_tokens: true,
Nicolas Patry's avatar
Nicolas Patry committed
1046
1047
                parameters: GenerateParameters {
                    top_n_tokens: Some(5),
1048
                    max_new_tokens: Some(5),
Nicolas Patry's avatar
Nicolas Patry committed
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
                    ..default_parameters()
                },
            })
            .await
        {
            Err(ValidationError::TopNTokens(4, 5)) => (),
            _ => panic!("Unexpected top_n_tokens"),
        }

        validation
            .validate(GenerateRequest {
                inputs: "Hello".to_string(),
1061
                add_special_tokens: true,
Nicolas Patry's avatar
Nicolas Patry committed
1062
1063
                parameters: GenerateParameters {
                    top_n_tokens: Some(4),
1064
                    max_new_tokens: Some(5),
Nicolas Patry's avatar
Nicolas Patry committed
1065
1066
1067
1068
1069
1070
1071
1072
1073
                    ..default_parameters()
                },
            })
            .await
            .unwrap();

        validation
            .validate(GenerateRequest {
                inputs: "Hello".to_string(),
1074
                add_special_tokens: true,
Nicolas Patry's avatar
Nicolas Patry committed
1075
1076
                parameters: GenerateParameters {
                    top_n_tokens: Some(0),
1077
                    max_new_tokens: Some(5),
Nicolas Patry's avatar
Nicolas Patry committed
1078
1079
1080
1081
1082
1083
1084
1085
1086
                    ..default_parameters()
                },
            })
            .await
            .unwrap();

        let valid_request = validation
            .validate(GenerateRequest {
                inputs: "Hello".to_string(),
1087
                add_special_tokens: true,
Nicolas Patry's avatar
Nicolas Patry committed
1088
1089
                parameters: GenerateParameters {
                    top_n_tokens: None,
1090
                    max_new_tokens: Some(5),
Nicolas Patry's avatar
Nicolas Patry committed
1091
1092
1093
1094
1095
1096
1097
1098
                    ..default_parameters()
                },
            })
            .await
            .unwrap();

        assert_eq!(valid_request.top_n_tokens, 0);
    }
1099
1100
1101
1102
1103
1104
1105

    static PIXEL_GIF: &str = "R0lGODdhAQABAIEAAP///wAAAAAAAAAAACwAAAAAAQABAAAIBAABBAQAOw==";

    #[tokio::test]
    async fn test_prepare_input_chunks() {
        let pixel_data = STANDARD.decode(PIXEL_GIF).unwrap();

1106
        let tokenizer = get_tokenizer();
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123

        let max_best_of = 2;
        let max_stop_sequence = 3;
        let max_top_n_tokens = 4;
        let max_input_length = 5;
        let max_total_tokens = 6;
        let disable_grammar_support = true;
        let workers = 1;
        let config = Config::Paligemma(Paligemma {
            text_config: PaliTextConfig {
                num_image_tokens: 1,
            },
        });
        let validation = Validation::new(
            workers,
            tokenizer,
            Some(config),
1124
            None,
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
            max_best_of,
            max_stop_sequence,
            max_top_n_tokens,
            max_input_length,
            max_total_tokens,
            disable_grammar_support,
        );

        let chunks = match validation
            .tokenize(
                format!("test![](data:image/gif;base64,{})", PIXEL_GIF),
1136
                true,
1137
1138
1139
1140
                None,
            )
            .await
        {
1141
            Ok((_encoding, chunks)) => chunks,
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
            _ => panic!("Unexpected tokenization failure"),
        };

        assert!(
            chunks
                == vec![
                    Chunk::Text("test".to_string()).into(),
                    Chunk::Image(Image {
                        data: pixel_data.clone(),
                        mimetype: "image/gif".to_string()
                    })
                    .into()
                ],
            "Failed to process images",
        );
    }
1158
1159
1160
1161
1162

    #[tokio::test]
    async fn test_idefics2_correct_n_fake_tokens() {
        let pixel_data = STANDARD.decode(PIXEL_GIF).unwrap();

1163
        let tokenizer = get_tokenizer();
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195

        let max_best_of = 2;
        let max_stop_sequence = 3;
        let max_top_n_tokens = 4;
        let max_input_length = 5;
        let max_total_tokens = 6;
        let disable_grammar_support = true;
        let workers = 1;
        let config = Config::Idefics2(Idefics2 {});
        let validation = Validation::new(
            workers,
            tokenizer,
            Some(config),
            Some(HubPreprocessorConfig::Idefics2Processor(
                Idefics2Preprocessor {
                    do_image_splitting: true,
                },
            )),
            max_best_of,
            max_stop_sequence,
            max_top_n_tokens,
            max_input_length,
            max_total_tokens,
            disable_grammar_support,
        );

        let (encoding, chunks) = match validation
            .tokenize(
                format!(
                    "test![](data:image/gif;base64,{})![](data:image/gif;base64,{})",
                    PIXEL_GIF, PIXEL_GIF
                ),
1196
                true,
1197
1198
1199
1200
                None,
            )
            .await
        {
1201
            Ok((encoding, chunks)) => (encoding, chunks),
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
            _ => panic!("Unexpected tokenization failure"),
        };

        assert!(
            chunks
                == vec![
                    Chunk::Text("test".to_string()).into(),
                    Chunk::Image(Image {
                        data: pixel_data.clone(),
                        mimetype: "image/gif".to_string()
                    })
                    .into(),
                    Chunk::Image(Image {
                        data: pixel_data.clone(),
                        mimetype: "image/gif".to_string()
                    })
                    .into()
                ],
            "Failed to process images",
        );

        // Verify the number of fake tokens:
        //
        // - Two images surrounded/separated by a fake token = 3.
        // - Both are split in 5 subimages, separated by a fake token: 2 * 4
        //
        // Fake tokens get split up by the testing tokenizer, but we don't care.
        assert_eq!(
            encoding
                .get_tokens()
                .iter()
                .filter(|t| *t == "fake")
                .count(),
            11
        );
    }
1238
}