lib.rs 55.8 KB
Newer Older
1
/// Text Generation Inference Webserver
OlivierDehaene's avatar
OlivierDehaene committed
2
pub mod config;
Nicolas Patry's avatar
Nicolas Patry committed
3
pub mod infer;
Olivier Dehaene's avatar
Olivier Dehaene committed
4
pub mod server;
Nicolas Patry's avatar
Nicolas Patry committed
5
pub mod validation;
Olivier Dehaene's avatar
Olivier Dehaene committed
6

7
8
#[cfg(feature = "kserve")]
mod kserve;
Nicolas Patry's avatar
Nicolas Patry committed
9
pub mod logging;
10

11
mod sagemaker;
12
pub mod usage_stats;
Nicolas Patry's avatar
Nicolas Patry committed
13
mod vertex;
14

15
use crate::infer::tool_grammar::ToolGrammar;
Nicolas Patry's avatar
Nicolas Patry committed
16
use crate::infer::{Infer, InferError};
17
18
use pyo3::prelude::*;
use pyo3::types::IntoPyDict;
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
19
use serde::{Deserialize, Serialize};
20
use tokenizers::Encoding;
Nicolas Patry's avatar
Nicolas Patry committed
21
use tracing::warn;
22
use utoipa::ToSchema;
Olivier Dehaene's avatar
Olivier Dehaene committed
23
use validation::Validation;
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
24

25
#[allow(clippy::large_enum_variant)]
26
27
28
29
30
#[derive(Clone)]
pub enum Tokenizer {
    Python {
        tokenizer_name: String,
        revision: Option<String>,
31
        trust_remote_code: bool,
32
33
34
35
36
37
38
39
40
41
42
    },
    Rust(tokenizers::Tokenizer),
}

pub struct PyTokenizer<'a>(pyo3::Bound<'a, pyo3::PyAny>);

impl<'a> PyTokenizer<'a> {
    fn from_py(
        py: Python<'a>,
        tokenizer_name: String,
        revision: Option<String>,
43
        trust_remote_code: bool,
44
45
46
47
48
49
    ) -> PyResult<PyTokenizer<'a>> {
        let transformers = py.import_bound("transformers")?;
        let auto = transformers.getattr("AutoTokenizer")?;
        let from_pretrained = auto.getattr("from_pretrained")?;
        let args = (tokenizer_name,);
        let kwargs = if let Some(rev) = &revision {
50
51
52
53
54
            [
                ("revision", rev.to_string().into_py(py)),
                ("trust_remote_code", trust_remote_code.into_py(py)),
            ]
            .into_py_dict_bound(py)
55
        } else {
56
            [("trust_remote_code", trust_remote_code.into_py(py))].into_py_dict_bound(py)
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
        };
        let tokenizer = from_pretrained.call(args, Some(&kwargs))?;
        tracing::info!("Loaded a python tokenizer");
        Ok(PyTokenizer(tokenizer))
    }
}

trait TokenizerTrait {
    fn encode_trait(
        &self,
        query: String,
        add_special_tokens: bool,
    ) -> Result<tokenizers::Encoding, Box<dyn std::error::Error + Send + Sync>>;
}

impl TokenizerTrait for tokenizers::Tokenizer {
    fn encode_trait(
        &self,
        query: String,
        add_special_tokens: bool,
    ) -> Result<tokenizers::Encoding, Box<dyn std::error::Error + Send + Sync>> {
        self.encode(query, add_special_tokens)
    }
}

impl<'a> TokenizerTrait for PyTokenizer<'a> {
    fn encode_trait(
        &self,
        query: String,
        add_special_tokens: bool,
    ) -> Result<tokenizers::Encoding, Box<dyn std::error::Error + Send + Sync>> {
        let py = self.0.py();
        let kwargs = [
            ("text", query.into_py(py)),
            ("add_special_tokens", add_special_tokens.into_py(py)),
        ]
        .into_py_dict_bound(py);
        let encode = self.0.getattr("encode")?;
        let input_ids: Vec<u32> = encode.call((), Some(&kwargs))?.extract()?;
        Ok(Encoding::new(
            input_ids,
            vec![],                           // type ids
            vec![],                           // tokens (strings)
            vec![],                           // words
            vec![],                           // offsets
            vec![],                           // special_tokens_mask
            vec![],                           // attention_mask
            vec![],                           // overflowing
            std::collections::HashMap::new(), //sequence_ranges
        ))
    }
}

110
111
/// Hub type
#[derive(Clone, Debug, Deserialize)]
112
pub struct HubModelInfo {
113
114
115
116
117
118
    #[serde(rename(deserialize = "id"))]
    pub model_id: String,
    pub sha: Option<String>,
    pub pipeline_tag: Option<String>,
}

119
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
120
121
122
123
124
pub struct ChatTemplate {
    name: String,
    template: String,
}

125
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
126
127
128
129
130
131
#[serde(untagged)]
pub enum ChatTemplateVersions {
    Single(String),
    Multiple(Vec<ChatTemplate>),
}

132
133
use std::path::Path;

134
#[derive(Debug, Clone, Serialize, Deserialize, Default)]
135
pub struct HubTokenizerConfig {
136
    pub chat_template: Option<ChatTemplateVersions>,
137
    pub completion_template: Option<String>,
138
139
    pub bos_token: Option<TokenizerConfigToken>,
    pub eos_token: Option<TokenizerConfigToken>,
140
141
142
    pub tokenizer_class: Option<String>,
    pub add_bos_token: Option<bool>,
    pub add_eos_token: Option<bool>,
143
144
145
}

impl HubTokenizerConfig {
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
    pub fn from_file<P: AsRef<Path>>(filename: P) -> Option<Self> {
        std::fs::read_to_string(filename)
            .ok()
            .and_then(|content| serde_json::from_str(&content).ok())
    }
}

#[derive(Debug, Clone, Deserialize, Serialize, PartialEq)]
#[serde(untagged)]
pub enum TokenizerConfigToken {
    String(String),
    Object { content: String },
}

impl TokenizerConfigToken {
    pub fn as_str(&self) -> &str {
        match self {
            TokenizerConfigToken::String(s) => s,
            TokenizerConfigToken::Object { content } => content,
        }
166
167
168
    }
}

169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(tag = "processor_class")]
pub enum HubPreprocessorConfig {
    Idefics2Processor(Idefics2Preprocessor),
}

impl HubPreprocessorConfig {
    pub fn from_file<P: AsRef<std::path::Path>>(filename: P) -> Option<Self> {
        let content = std::fs::read_to_string(filename).ok()?;
        serde_json::from_str(&content).ok()
    }
}

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Idefics2Preprocessor {
    #[serde(default)]
    do_image_splitting: bool,
}

drbh's avatar
drbh committed
188
189
190
191
192
193
194
195
#[derive(Debug, Clone, Deserialize, Default)]
pub struct HubProcessorConfig {
    pub chat_template: Option<ChatTemplateVersions>,
    pub image_seq_len: usize,
    pub processor_class: Option<String>,
}

impl HubProcessorConfig {
196
197
198
199
    pub fn from_file<P: AsRef<Path>>(filename: P) -> Option<Self> {
        std::fs::read_to_string(filename)
            .ok()
            .and_then(|content| serde_json::from_str(&content).ok())
drbh's avatar
drbh committed
200
201
202
    }
}

203
#[derive(Clone, Debug, Deserialize, ToSchema, Serialize)]
Nicolas Patry's avatar
Nicolas Patry committed
204
#[cfg_attr(test, derive(PartialEq))]
drbh's avatar
drbh committed
205
206
#[serde(tag = "type", content = "value")]
pub(crate) enum GrammarType {
207
208
209
210
211
    /// A string that represents a [JSON Schema](https://json-schema.org/).
    ///
    /// JSON Schema is a declarative language that allows to annotate JSON documents
    /// with types and descriptions.
    #[serde(rename = "json")]
drbh's avatar
drbh committed
212
    #[serde(alias = "json_object")]
213
214
    #[schema(example = json ! ({"properties": {"location":{"type": "string"}}}))]
    Json(serde_json::Value),
drbh's avatar
drbh committed
215
216
217
218
    #[serde(rename = "regex")]
    Regex(String),
}

219
220
#[derive(Clone, Debug, Serialize, ToSchema)]
pub struct Info {
221
    /// Model info
222
223
224
225
    #[schema(example = "bigscience/blomm-560m")]
    pub model_id: String,
    #[schema(nullable = true, example = "e985a63cdc139290c5f700ff1929f0b5942cced2")]
    pub model_sha: Option<String>,
Nicolas Patry's avatar
Nicolas Patry committed
226
227
228
229
    // #[schema(example = "torch.float16")]
    // pub model_dtype: String,
    // #[schema(example = "cuda")]
    // pub model_device_type: String,
230
231
    #[schema(nullable = true, example = "text-generation")]
    pub model_pipeline_tag: Option<String>,
Nicolas Patry's avatar
Nicolas Patry committed
232

233
234
235
236
237
238
239
240
    /// Router Parameters
    #[schema(example = "128")]
    pub max_concurrent_requests: usize,
    #[schema(example = "2")]
    pub max_best_of: usize,
    #[schema(example = "4")]
    pub max_stop_sequences: usize,
    #[schema(example = "1024")]
OlivierDehaene's avatar
OlivierDehaene committed
241
    pub max_input_tokens: usize,
242
243
244
245
    #[schema(example = "2048")]
    pub max_total_tokens: usize,
    #[schema(example = "2")]
    pub validation_workers: usize,
246
247
    #[schema(example = "32")]
    pub max_client_batch_size: usize,
Nicolas Patry's avatar
Nicolas Patry committed
248

249
    /// Router Info
250
251
    #[schema(example = "text-generation-router")]
    pub router: &'static str,
252
253
254
255
    #[schema(example = "0.5.0")]
    pub version: &'static str,
    #[schema(nullable = true, example = "null")]
    pub sha: Option<&'static str>,
256
257
    #[schema(nullable = true, example = "null")]
    pub docker_label: Option<&'static str>,
258
259
}

drbh's avatar
drbh committed
260
#[derive(Clone, Debug, Deserialize, ToSchema, Default)]
Nicolas Patry's avatar
Nicolas Patry committed
261
#[cfg_attr(test, derive(PartialEq))]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
262
pub(crate) struct GenerateParameters {
263
    /// Generate best_of sequences and return the one if the highest token logprobs.
264
265
266
    #[serde(default)]
    #[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 1)]
    pub best_of: Option<usize>,
267
268

    /// The value used to module the logits distribution.
269
270
271
272
273
274
275
276
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        nullable = true,
        default = "null",
        example = 0.5
    )]
    pub temperature: Option<f32>,
277
278
279

    /// The parameter for repetition penalty. 1.0 means no penalty.
    /// See [this paper](https://arxiv.org/pdf/1909.05858.pdf) for more details.
280
281
282
283
284
285
286
287
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        nullable = true,
        default = "null",
        example = 1.03
    )]
    pub repetition_penalty: Option<f32>,
288
289
290
291

    /// The parameter for frequency penalty. 1.0 means no penalty
    /// Penalize new tokens based on their existing frequency in the text so far,
    /// decreasing the model's likelihood to repeat the same line verbatim.
292
    #[serde(default)]
293
294
295
296
297
298
299
    #[schema(
        exclusive_minimum = -2.0,
        nullable = true,
        default = "null",
        example = 0.1
    )]
    pub frequency_penalty: Option<f32>,
300
301

    /// The number of highest probability vocabulary tokens to keep for top-k-filtering.
302
    #[serde(default)]
303
304
    #[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 10)]
    pub top_k: Option<i32>,
305
306

    /// Top-p value for nucleus sampling.
307
308
309
310
311
312
313
314
315
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        maximum = 1.0,
        nullable = true,
        default = "null",
        example = 0.95
    )]
    pub top_p: Option<f32>,
316
317
318

    /// Typical Decoding mass
    /// See [Typical Decoding for Natural Language Generation](https://arxiv.org/abs/2202.00666) for more information.
319
    #[serde(default)]
320
321
322
323
324
325
326
327
    #[schema(
        exclusive_minimum = 0.0,
        maximum = 1.0,
        nullable = true,
        default = "null",
        example = 0.95
    )]
    pub typical_p: Option<f32>,
328
329

    /// Activate logits sampling.
330
    #[serde(default)]
331
    #[schema(default = "false", example = true)]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
332
    pub do_sample: bool,
333
334

    /// Maximum number of tokens to generate.
335
336
    #[serde(default)]
    #[schema(nullable = true, default = "1024", example = "20")]
337
    pub max_new_tokens: Option<u32>,
338
339

    /// Whether to prepend the prompt to the generated text
OlivierDehaene's avatar
OlivierDehaene committed
340
    #[serde(default)]
341
    #[schema(nullable = true, default = "null", example = false)]
342
    pub return_full_text: Option<bool>,
343
344

    /// Stop generating tokens if a member of `stop` is generated.
345
    #[serde(default)]
346
    #[schema(inline, max_items = 4, example = json ! (["photographer"]))]
347
    pub stop: Vec<String>,
348
349

    /// Truncate inputs tokens to the given size.
OlivierDehaene's avatar
OlivierDehaene committed
350
    #[serde(default)]
351
    #[schema(nullable = true, default = "null", example = "null")]
352
    pub truncate: Option<usize>,
353
354

    /// Watermarking with [A Watermark for Large Language Models](https://arxiv.org/abs/2301.10226).
355
    #[serde(default)]
356
357
    #[schema(default = "false", example = true)]
    pub watermark: bool,
358
359

    /// Whether to return generation details.
360
    #[serde(default)]
361
    #[schema(default = "true")]
OlivierDehaene's avatar
OlivierDehaene committed
362
    pub details: bool,
363
364

    /// Whether to return decoder input token logprobs and ids.
365
    #[serde(default)]
366
    #[schema(default = "false")]
367
    pub decoder_input_details: bool,
368
369

    /// Random sampling seed.
370
    #[serde(default)]
371
372
373
374
375
376
    #[schema(
        exclusive_minimum = 0,
        nullable = true,
        default = "null",
        example = "null"
    )]
377
    pub seed: Option<u64>,
378
379

    /// The number of highest probability vocabulary tokens to keep for top-n-filtering.
Nicolas Patry's avatar
Nicolas Patry committed
380
381
382
    #[serde(default)]
    #[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 5)]
    pub top_n_tokens: Option<u32>,
383
384

    /// Grammar constraints for the generation.
drbh's avatar
drbh committed
385
    #[serde(default)]
386
    #[schema(nullable = true, default = "null", example = "null")]
drbh's avatar
drbh committed
387
    pub grammar: Option<GrammarType>,
drbh's avatar
drbh committed
388
389
390
391
392

    /// Lora adapter id
    #[serde(default)]
    #[schema(nullable = true, default = "null", example = "null")]
    pub adapter_id: Option<String>,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
393
394
395
396
}

fn default_parameters() -> GenerateParameters {
    GenerateParameters {
397
        best_of: None,
398
399
        temperature: None,
        repetition_penalty: None,
400
        frequency_penalty: None,
401
402
        top_k: None,
        top_p: None,
403
        typical_p: None,
404
        do_sample: true,
405
        max_new_tokens: None,
406
        return_full_text: None,
407
        stop: Vec::new(),
408
        truncate: None,
409
        watermark: false,
OlivierDehaene's avatar
OlivierDehaene committed
410
        details: false,
411
        decoder_input_details: false,
412
        seed: None,
Nicolas Patry's avatar
Nicolas Patry committed
413
        top_n_tokens: None,
drbh's avatar
drbh committed
414
        grammar: None,
drbh's avatar
drbh committed
415
        adapter_id: None,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
416
417
418
    }
}

419
420
421
422
423
424
425
426
427
428
429
430
431
#[derive(Clone, Deserialize, Serialize, ToSchema, Debug)]
#[serde(try_from = "PromptDeserializer")]
pub struct Prompt(pub Vec<String>);

#[derive(Deserialize)]
#[serde(untagged)]
enum PromptDeserializer {
    Single(String),
    Multiple(Vec<String>),
}

impl TryFrom<PromptDeserializer> for Prompt {
    type Error = String;
432

433
    fn try_from(value: PromptDeserializer) -> Result<Self, Self::Error> {
434
        match value {
435
436
437
438
439
440
441
442
443
444
445
            PromptDeserializer::Single(s) => Ok(Prompt(vec![s])),
            PromptDeserializer::Multiple(v) => {
                if v.is_empty() {
                    Err(
                        "Empty array detected. Do not use an empty array for the prompt."
                            .to_string(),
                    )
                } else {
                    Ok(Prompt(v))
                }
            }
446
447
448
449
        }
    }
}

450
451
452
453
454
#[derive(Clone, Deserialize, Serialize, ToSchema, Debug)]
pub struct CompletionRequest {
    /// UNUSED
    #[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
    /// ID of the model to use. See the model endpoint compatibility table for details on which models work with the Chat API.
455
    pub model: Option<String>,
456
457
458

    /// The prompt to generate completions for.
    #[schema(example = "What is Deep Learning?")]
459
    pub prompt: Prompt,
460
461
462

    /// The maximum number of tokens that can be generated in the chat completion.
    #[serde(default)]
463
    #[schema(default = "1024", example = "32")]
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
    pub max_tokens: Option<u32>,

    /// What sampling temperature to use, between 0 and 2. Higher values like 0.8 will make the output more random, while
    /// lower values like 0.2 will make it more focused and deterministic. We generally recommend altering this or `top_p` but not both.
    #[serde(default)]
    #[schema(nullable = true, example = 1.0)]
    pub temperature: Option<f32>,

    /// An alternative to sampling with temperature, called nucleus sampling, where the model considers the results of the
    /// tokens with top_p probability mass. So 0.1 means only the tokens comprising the top 10% probability mass are considered.
    #[serde(default)]
    #[schema(nullable = true, example = 0.95)]
    pub top_p: Option<f32>,

    #[serde(default = "bool::default")]
    pub stream: bool,

    #[schema(nullable = true, example = 42)]
    pub seed: Option<u64>,

    /// The text to append to the prompt. This is useful for completing sentences or generating a paragraph of text.
    /// please see the completion_template field in the model's tokenizer_config.json file for completion template.
    #[serde(default)]
    pub suffix: Option<String>,

    #[serde(default)]
    pub repetition_penalty: Option<f32>,

    /// Number between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far,
    /// decreasing the model's likelihood to repeat the same line verbatim.
    #[serde(default)]
    #[schema(example = "1.0")]
    pub frequency_penalty: Option<f32>,
497
498
499
500
501

    /// Up to 4 sequences where the API will stop generating further tokens.
    #[serde(default)]
    #[schema(nullable = true, example = "null")]
    pub stop: Option<Vec<String>>,
502
503
}

504
505
506
507
508
509
510
511
512
#[derive(Clone, Serialize, ToSchema)]
#[serde(tag = "object")]
enum Completion {
    #[serde(rename = "text_completion")]
    Chunk(Chunk),
    #[serde(rename = "text_completion")]
    Final(CompletionFinal),
}

513
#[derive(Clone, Deserialize, Serialize, ToSchema, Default)]
514
pub(crate) struct CompletionFinal {
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
    pub id: String,
    #[schema(example = "1706270835")]
    pub created: u64,
    #[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
    pub model: String,
    pub system_fingerprint: String,
    pub choices: Vec<CompletionComplete>,
    pub usage: Usage,
}

#[derive(Clone, Deserialize, Serialize, ToSchema)]
pub(crate) struct CompletionComplete {
    pub index: u32,
    pub text: String,
    pub logprobs: Option<Vec<f32>>,
    pub finish_reason: String,
}

533
534
535
536
537
538
539
540
541
#[derive(Clone, Deserialize, Serialize, ToSchema)]
pub(crate) struct Chunk {
    pub id: String,
    pub created: u64,
    pub choices: Vec<CompletionComplete>,
    pub model: String,
    pub system_fingerprint: String,
}

542
#[derive(Clone, Deserialize, Serialize, ToSchema)]
543
544
pub(crate) struct ChatCompletion {
    pub id: String,
545
    #[schema(example = "1706270835")]
546
    pub created: u64,
547
    #[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
548
549
550
551
552
553
    pub model: String,
    pub system_fingerprint: String,
    pub choices: Vec<ChatCompletionComplete>,
    pub usage: Usage,
}

554
#[derive(Clone, Deserialize, Serialize, ToSchema)]
555
556
pub(crate) struct ChatCompletionComplete {
    pub index: u32,
Nicolas Patry's avatar
Nicolas Patry committed
557
    pub message: OutputMessage,
558
    pub logprobs: Option<ChatCompletionLogprobs>,
559
560
561
    pub finish_reason: String,
}

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
#[derive(Clone, Deserialize, Serialize, ToSchema)]
pub(crate) struct ChatCompletionLogprobs {
    content: Vec<ChatCompletionLogprob>,
}

impl From<(Token, Vec<Token>)> for ChatCompletionLogprobs {
    fn from(value: (Token, Vec<Token>)) -> Self {
        let (token, top_tokens) = value;

        Self {
            content: vec![ChatCompletionLogprob {
                token: token.text,
                logprob: token.logprob,
                top_logprobs: top_tokens
                    .into_iter()
                    .map(|t| ChatCompletionTopLogprob {
                        token: t.text,
                        logprob: t.logprob,
                    })
                    .collect(),
            }],
        }
    }
}

impl From<(Vec<Token>, Vec<Vec<Token>>)> for ChatCompletionLogprobs {
    fn from(value: (Vec<Token>, Vec<Vec<Token>>)) -> Self {
        let (tokens, top_tokens) = value;
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604

        // Create an iterator that produces None for top_tokens once it's exhausted
        let top_tokens_iter = top_tokens
            .into_iter()
            .map(Some)
            .chain(std::iter::repeat(None));

        let content = tokens
            .into_iter()
            .zip(top_tokens_iter)
            .map(|(t, top_t_option)| ChatCompletionLogprob {
                token: t.text,
                logprob: t.logprob,
                top_logprobs: match top_t_option {
                    Some(top_t) => top_t
605
606
607
608
609
610
                        .into_iter()
                        .map(|t| ChatCompletionTopLogprob {
                            token: t.text,
                            logprob: t.logprob,
                        })
                        .collect(),
611
612
613
614
615
616
                    None => vec![], // Handle the case where there are no top tokens
                },
            })
            .collect();

        Self { content }
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
    }
}

#[derive(Clone, Deserialize, Serialize, ToSchema)]
pub(crate) struct ChatCompletionLogprob {
    token: String,
    logprob: f32,
    top_logprobs: Vec<ChatCompletionTopLogprob>,
}

#[derive(Clone, Deserialize, Serialize, ToSchema)]
pub(crate) struct ChatCompletionTopLogprob {
    token: String,
    logprob: f32,
}

633
#[derive(Clone, Deserialize, Serialize, ToSchema, Default)]
634
635
636
637
638
639
pub(crate) struct Usage {
    pub prompt_tokens: u32,
    pub completion_tokens: u32,
    pub total_tokens: u32,
}

640
641
642
643
644
645
646
647
648
#[derive(Clone, Serialize, ToSchema)]
#[serde(tag = "object")]
enum CompletionType {
    #[serde(rename = "chat.completion.chunk")]
    ChatCompletionChunk(ChatCompletionChunk),
    #[serde(rename = "chat.completion")]
    ChatCompletion(ChatCompletion),
}

649
impl ChatCompletion {
Nicolas Patry's avatar
Nicolas Patry committed
650
    #[allow(clippy::too_many_arguments)]
651
652
653
    pub(crate) fn new(
        model: String,
        system_fingerprint: String,
drbh's avatar
drbh committed
654
        output: Option<String>,
655
656
657
        created: u64,
        details: Details,
        return_logprobs: bool,
658
        tool_calls: Option<Vec<ToolCall>>,
Nicolas Patry's avatar
Nicolas Patry committed
659
        prompt_tokens: u32,
660
    ) -> Self {
Nicolas Patry's avatar
Nicolas Patry committed
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
        let message = match (output, tool_calls) {
            (Some(content), None) => OutputMessage::ChatMessage(TextMessage {
                role: "assistant".into(),
                content,
            }),
            (None, Some(tool_calls)) => OutputMessage::ToolCall(ToolCallMessage {
                role: "assistant".to_string(),
                tool_calls,
            }),
            (Some(output), Some(_)) => {
                warn!("Received both chat and tool call");
                OutputMessage::ChatMessage(TextMessage {
                    role: "assistant".into(),
                    content: output,
                })
            }
            (None, None) => {
                warn!("Didn't receive an answer");
                OutputMessage::ChatMessage(TextMessage {
                    role: "assistant".into(),
                    content: "".to_string(),
                })
            }
        };
685
686
687
688
689
690
691
        Self {
            id: String::new(),
            created,
            model,
            system_fingerprint,
            choices: vec![ChatCompletionComplete {
                index: 0,
Nicolas Patry's avatar
Nicolas Patry committed
692
                message,
693
                logprobs: return_logprobs
694
                    .then(|| ChatCompletionLogprobs::from((details.tokens, details.top_tokens))),
695
                finish_reason: details.finish_reason.format(true),
696
697
            }],
            usage: Usage {
Nicolas Patry's avatar
Nicolas Patry committed
698
                prompt_tokens,
699
                completion_tokens: details.generated_tokens,
Nicolas Patry's avatar
Nicolas Patry committed
700
                total_tokens: prompt_tokens + details.generated_tokens,
701
702
703
704
            },
        }
    }
}
705
#[derive(Clone, Serialize, ToSchema)]
706
707
pub(crate) struct ChatCompletionChunk {
    pub id: String,
708
    #[schema(example = "1706270978")]
709
    pub created: u64,
710
    #[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
711
712
713
    pub model: String,
    pub system_fingerprint: String,
    pub choices: Vec<ChatCompletionChoice>,
Nicolas Patry's avatar
Nicolas Patry committed
714
    pub usage: Option<Usage>,
715
716
}

717
#[derive(Clone, Serialize, ToSchema)]
718
719
720
pub(crate) struct ChatCompletionChoice {
    pub index: u32,
    pub delta: ChatCompletionDelta,
721
    pub logprobs: Option<ChatCompletionLogprobs>,
722
723
724
    pub finish_reason: Option<String>,
}

Nicolas Patry's avatar
Nicolas Patry committed
725
726
727
728
729
730
731
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
pub struct ToolCallDelta {
    #[schema(example = "assistant")]
    role: String,
    tool_calls: DeltaToolCall,
}

732
733
#[derive(Clone, Debug, Serialize, ToSchema)]
#[serde(untagged)]
Nicolas Patry's avatar
Nicolas Patry committed
734
735
736
enum ChatCompletionDelta {
    Chat(TextMessage),
    Tool(ToolCallDelta),
drbh's avatar
drbh committed
737
738
}

Nicolas Patry's avatar
Nicolas Patry committed
739
#[derive(Clone, Deserialize, Serialize, ToSchema, Debug, PartialEq)]
drbh's avatar
drbh committed
740
741
742
743
744
745
746
pub(crate) struct DeltaToolCall {
    pub index: u32,
    pub id: String,
    pub r#type: String,
    pub function: Function,
}

Nicolas Patry's avatar
Nicolas Patry committed
747
#[derive(Clone, Deserialize, Serialize, ToSchema, Debug, PartialEq)]
drbh's avatar
drbh committed
748
749
750
pub(crate) struct Function {
    pub name: Option<String>,
    pub arguments: String,
751
752
}

drbh's avatar
drbh committed
753
#[allow(clippy::too_many_arguments)]
754
755
756
757
impl ChatCompletionChunk {
    pub(crate) fn new(
        model: String,
        system_fingerprint: String,
drbh's avatar
drbh committed
758
759
        delta: Option<String>,
        tool_calls: Option<Vec<String>>,
760
        created: u64,
761
        logprobs: Option<ChatCompletionLogprobs>,
762
        finish_reason: Option<String>,
Nicolas Patry's avatar
Nicolas Patry committed
763
        usage: Option<Usage>,
764
    ) -> Self {
765
        let delta = match (delta, tool_calls) {
Nicolas Patry's avatar
Nicolas Patry committed
766
767
768
769
770
771
772
            (Some(delta), _) => ChatCompletionDelta::Chat(TextMessage {
                role: "assistant".to_string(),
                content: delta,
            }),
            (None, Some(tool_calls)) => ChatCompletionDelta::Tool(ToolCallDelta {
                role: "assistant".to_string(),
                tool_calls: DeltaToolCall {
773
774
775
776
777
778
779
                    index: 0,
                    id: String::new(),
                    r#type: "function".to_string(),
                    function: Function {
                        name: None,
                        arguments: tool_calls[0].to_string(),
                    },
Nicolas Patry's avatar
Nicolas Patry committed
780
781
782
783
784
785
                },
            }),
            (None, None) => ChatCompletionDelta::Chat(TextMessage {
                role: "assistant".to_string(),
                content: "".to_string(),
            }),
786
        };
787
788
789
790
791
792
        Self {
            id: String::new(),
            created,
            model,
            system_fingerprint,
            choices: vec![ChatCompletionChoice {
793
                index: 0,
794
                delta,
795
796
797
                logprobs,
                finish_reason,
            }],
Nicolas Patry's avatar
Nicolas Patry committed
798
            usage,
799
800
801
802
803
        }
    }
}

#[derive(Clone, Deserialize, ToSchema, Serialize)]
Nicolas Patry's avatar
Nicolas Patry committed
804
#[cfg_attr(test, derive(Debug, PartialEq, Default))]
805
pub(crate) struct ChatRequest {
806
    #[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
drbh's avatar
drbh committed
807
    /// [UNUSED] ID of the model to use. See the model endpoint compatibility table for details on which models work with the Chat API.
808
    pub model: Option<String>,
drbh's avatar
drbh committed
809

810
    /// A list of messages comprising the conversation so far.
drbh's avatar
drbh committed
811
    #[schema(example = "[{\"role\": \"user\", \"content\": \"What is Deep Learning?\"}]")]
812
813
814
815
816
    pub messages: Vec<Message>,

    /// Number between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far,
    /// decreasing the model's likelihood to repeat the same line verbatim.
    #[serde(default)]
817
    #[schema(example = "1.0")]
818
819
820
821
822
823
824
825
826
827
828
829
830
831
    pub frequency_penalty: Option<f32>,

    /// UNUSED
    /// Modify the likelihood of specified tokens appearing in the completion. Accepts a JSON object that maps tokens
    /// (specified by their token ID in the tokenizer) to an associated bias value from -100 to 100. Mathematically,
    /// the bias is added to the logits generated by the model prior to sampling. The exact effect will vary per model,
    /// but values between -1 and 1 should decrease or increase likelihood of selection; values like -100 or 100 should
    /// result in a ban or exclusive selection of the relevant token.
    #[serde(default)]
    pub logit_bias: Option<Vec<f32>>,

    /// Whether to return log probabilities of the output tokens or not. If true, returns the log probabilities of each
    /// output token returned in the content of message.
    #[serde(default)]
832
    #[schema(example = "false")]
833
834
835
836
837
    pub logprobs: Option<bool>,

    /// An integer between 0 and 5 specifying the number of most likely tokens to return at each token position, each with
    /// an associated log probability. logprobs must be set to true if this parameter is used.
    #[serde(default)]
838
    #[schema(example = "5")]
839
840
841
842
    pub top_logprobs: Option<u32>,

    /// The maximum number of tokens that can be generated in the chat completion.
    #[serde(default)]
843
    #[schema(default = "1024", example = "32")]
844
845
846
847
848
849
    pub max_tokens: Option<u32>,

    /// UNUSED
    /// How many chat completion choices to generate for each input message. Note that you will be charged based on the
    /// number of generated tokens across all of the choices. Keep n as 1 to minimize costs.
    #[serde(default)]
850
    #[schema(nullable = true, example = "2")]
851
852
853
854
855
    pub n: Option<u32>,

    /// Number between -2.0 and 2.0. Positive values penalize new tokens based on whether they appear in the text so far,
    /// increasing the model's likelihood to talk about new topics
    #[serde(default)]
856
    #[schema(nullable = true, example = 0.1)]
857
858
    pub presence_penalty: Option<f32>,

859
860
861
862
863
    /// Up to 4 sequences where the API will stop generating further tokens.
    #[serde(default)]
    #[schema(nullable = true, example = "null")]
    pub stop: Option<Vec<String>>,

864
865
866
867
868
    #[serde(default = "bool::default")]
    pub stream: bool,

    #[schema(nullable = true, example = 42)]
    pub seed: Option<u64>,
869
870
871
872
873
874

    /// What sampling temperature to use, between 0 and 2. Higher values like 0.8 will make the output more random, while
    /// lower values like 0.2 will make it more focused and deterministic.
    ///
    /// We generally recommend altering this or `top_p` but not both.
    #[serde(default)]
875
    #[schema(nullable = true, example = 1.0)]
876
877
878
879
880
    pub temperature: Option<f32>,

    /// An alternative to sampling with temperature, called nucleus sampling, where the model considers the results of the
    /// tokens with top_p probability mass. So 0.1 means only the tokens comprising the top 10% probability mass are considered.
    #[serde(default)]
881
    #[schema(nullable = true, example = 0.95)]
882
    pub top_p: Option<f32>,
drbh's avatar
drbh committed
883
884
885
886
887
888
889
890

    /// A list of tools the model may call. Currently, only functions are supported as a tool. Use this to provide a list of
    /// functions the model may generate JSON inputs for.
    #[serde(default)]
    #[schema(nullable = true, example = "null")]
    pub tools: Option<Vec<Tool>>,

    /// A prompt to be appended before the tools
drbh's avatar
drbh committed
891
    #[serde(default)]
drbh's avatar
drbh committed
892
893
    #[schema(
        nullable = true,
drbh's avatar
drbh committed
894
        example = "Given the functions available, please respond with a JSON for a function call with its proper arguments that best answers the given prompt. Respond in the format {name: function name, parameters: dictionary of argument name and its value}.Do not use variables."
drbh's avatar
drbh committed
895
896
897
898
899
    )]
    pub tool_prompt: Option<String>,

    /// A specific tool to use. If not provided, the model will default to use any of the tools provided in the tools parameter.
    #[serde(default)]
900
    #[schema(nullable = true, default = "auto", example = "auto")]
drbh's avatar
drbh committed
901
    pub tool_choice: ToolChoice,
drbh's avatar
drbh committed
902
903
904
905
906
907
908

    /// Response format constraints for the generation.
    ///
    /// NOTE: A request can use `response_format` OR `tools` but not both.
    #[serde(default)]
    #[schema(nullable = true, default = "null", example = "null")]
    pub response_format: Option<GrammarType>,
909

Nicolas Patry's avatar
Nicolas Patry committed
910
911
912
913
914
915
    /// Options for streaming response. Only set this when you set stream: true.
    #[serde(default)]
    #[schema(nullable = true, example = "null")]
    pub stream_options: Option<StreamOptions>,
}

Nicolas Patry's avatar
Nicolas Patry committed
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
impl ChatRequest {
    fn try_into_generate(self, infer: &Infer) -> Result<(GenerateRequest, bool), InferError> {
        let ChatRequest {
            model,
            max_tokens,
            messages,
            seed,
            stop,
            tools,
            tool_choice,
            tool_prompt,
            temperature,
            response_format,
            presence_penalty,
            frequency_penalty,
            top_p,
            top_logprobs,
            ..
        } = self;

        let repetition_penalty = presence_penalty.map(|x| x + 2.0);
937
        let max_new_tokens = max_tokens;
Nicolas Patry's avatar
Nicolas Patry committed
938
939
940
941
942
943
944
945
946
        let tool_prompt = tool_prompt
            .filter(|s| !s.is_empty())
            .unwrap_or_else(default_tool_prompt);
        let stop = stop.unwrap_or_default();
        // enable greedy only when temperature is 0
        let (do_sample, temperature) = match temperature {
            Some(temperature) if temperature == 0.0 => (false, None),
            other => (true, other),
        };
947
948
949
950
951
952
953
954
955

        if response_format.is_some() && tools.is_some() {
            return Err(InferError::ToolError(
                "Grammar and tools are mutually exclusive".into(),
            ));
        }

        let (inputs, grammar, using_tools) = match response_format {
            Some(format) => {
Lucain's avatar
Lucain committed
956
                let inputs = infer.apply_chat_template(messages, None)?;
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
                (inputs, Some(format), false)
            }
            None => {
                if let Some(tools) = tools {
                    match ToolGrammar::apply(tools, tool_choice)? {
                        Some((updated_tools, tool_schema)) => {
                            let grammar = GrammarType::Json(serde_json::json!(tool_schema));
                            let inputs: String = infer.apply_chat_template(
                                messages,
                                Some((updated_tools, tool_prompt)),
                            )?;
                            (inputs, Some(grammar), true)
                        }
                        None => {
                            // same as if no response_format or tools are set
Lucain's avatar
Lucain committed
972
                            let inputs = infer.apply_chat_template(messages, None)?;
973
974
975
976
977
                            (inputs, None, false)
                        }
                    }
                } else {
                    // if no response_format or tools are set simply apply the chat template to generate inputs
Lucain's avatar
Lucain committed
978
                    let inputs = infer.apply_chat_template(messages, None)?;
979
980
981
982
                    (inputs, None, false)
                }
            }
        };
Nicolas Patry's avatar
Nicolas Patry committed
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002

        Ok((
            GenerateRequest {
                inputs: inputs.to_string(),
                add_special_tokens: false,
                parameters: GenerateParameters {
                    best_of: None,
                    temperature,
                    repetition_penalty,
                    frequency_penalty,
                    top_k: None,
                    top_p,
                    typical_p: None,
                    do_sample,
                    max_new_tokens,
                    return_full_text: None,
                    stop,
                    truncate: None,
                    watermark: false,
                    details: true,
Nicolas Patry's avatar
Nicolas Patry committed
1003
                    decoder_input_details: false,
Nicolas Patry's avatar
Nicolas Patry committed
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
                    seed,
                    top_n_tokens: top_logprobs,
                    grammar,
                    adapter_id: model.filter(|m| *m != "tgi").map(String::from),
                },
            },
            using_tools,
        ))
    }
}

Nicolas Patry's avatar
Nicolas Patry committed
1015
#[derive(Clone, Deserialize, ToSchema, Serialize)]
Nicolas Patry's avatar
Nicolas Patry committed
1016
#[cfg_attr(test, derive(Debug, PartialEq))]
Nicolas Patry's avatar
Nicolas Patry committed
1017
1018
1019
1020
struct StreamOptions {
    /// If set, an additional chunk will be streamed before the data: [DONE] message. The usage field on this chunk shows the token usage statistics for the entire request, and the choices field will always be an empty array. All other chunks will also include a usage field, but with a null value.
    #[schema(example = "true")]
    include_usage: bool,
drbh's avatar
drbh committed
1021
1022
}

drbh's avatar
drbh committed
1023
1024
pub fn default_tool_prompt() -> String {
    "\nGiven the functions available, please respond with a JSON for a function call with its proper arguments that best answers the given prompt. Respond in the format {name: function name, parameters: dictionary of argument name and its value}.Do not use variables.\n".to_string()
drbh's avatar
drbh committed
1025
}
1026

1027
1028
1029
1030
1031
#[derive(Clone, Debug, Deserialize, PartialEq, Serialize)]
#[serde(tag = "type")]
pub enum TypedChoice {
    #[serde(rename = "function")]
    Function { function: FunctionName },
drbh's avatar
drbh committed
1032
1033
}

1034
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, ToSchema)]
1035
1036
1037
1038
pub struct FunctionName {
    pub name: String,
}

1039
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, ToSchema, Default)]
1040
#[serde(from = "ToolTypeDeserializer")]
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
#[serde(rename_all = "snake_case")]
/// <https://platform.openai.com/docs/guides/function-calling/configuring-function-calling-behavior-using-the-tool_choice-parameter>
pub enum ToolChoice {
    /// Means the model can pick between generating a message or calling one or more tools.
    #[default]
    Auto,
    /// Means the model will not call any tool and instead generates a message.
    #[serde(rename = "none")]
    NoTool,
    /// Means the model must call one or more tools.
    Required,
    /// Forces the model to call a specific tool. This structure aligns with the `OpenAI` API schema to force a specific tool.
    Function(FunctionName),
}
drbh's avatar
drbh committed
1055

1056
#[derive(Deserialize, ToSchema)]
1057
#[serde(untagged)]
1058
1059
1060
1061
1062
1063
1064
/// Controls which (if any) tool is called by the model.
/// - `none` means the model will not call any tool and instead generates a message.
/// - `auto` means the model can pick between generating a message or calling one or more tools.
/// - `required` means the model must call one or more tools.
/// - Specifying a particular tool via `{\"type\": \"function\", \"function\": {\"name\": \"my_function\"}}` forces the model to call that tool.
///
/// `none` is the default when no tools are present. `auto` is the default if tools are present."
1065
enum ToolTypeDeserializer {
1066
    /// None means `null` was passed in the JSON, and the default choice is applied based on the presence of tools.
1067
    Null,
1068
1069
1070

    /// `auto` means the model can pick between generating a message or calling one or more tools.
    #[schema(example = "auto")]
drbh's avatar
drbh committed
1071
    String(String),
1072
1073
1074
1075

    /// Specifying a particular tool forces the model to call that tool, with structured function details.
    #[schema(example = r#"{"type": "function", "function": {"name": "my_function"}}"#)]
    TypedChoice(TypedChoice),
1076
}
drbh's avatar
drbh committed
1077

1078
1079
impl From<ToolTypeDeserializer> for ToolChoice {
    fn from(value: ToolTypeDeserializer) -> Self {
drbh's avatar
drbh committed
1080
        match value {
1081
            ToolTypeDeserializer::Null => ToolChoice::Auto,
drbh's avatar
drbh committed
1082
            ToolTypeDeserializer::String(s) => match s.as_str() {
1083
1084
1085
1086
                "none" => ToolChoice::NoTool,
                "auto" => ToolChoice::Auto,
                "required" => ToolChoice::Required,
                _ => ToolChoice::Function(FunctionName { name: s }),
drbh's avatar
drbh committed
1087
            },
1088
1089
1090
            ToolTypeDeserializer::TypedChoice(TypedChoice::Function { function }) => {
                ToolChoice::Function(function)
            }
drbh's avatar
drbh committed
1091
1092
1093
1094
        }
    }
}

1095
#[derive(Debug, Deserialize, Serialize, ToSchema, PartialEq)]
drbh's avatar
drbh committed
1096
pub struct JsonSchemaTool {
drbh's avatar
drbh committed
1097
1098
1099
1100
1101
    #[serde(flatten)]
    functions_map: FunctionsMap,
    properties: Properties,
}

1102
#[derive(Debug, Serialize, Deserialize, PartialEq)]
drbh's avatar
drbh committed
1103
1104
1105
1106
1107
struct FunctionsMap {
    #[serde(rename = "$functions")]
    functions: std::collections::HashMap<String, serde_json::Value>,
}

1108
#[derive(Debug, Serialize, Deserialize, PartialEq)]
drbh's avatar
drbh committed
1109
1110
1111
1112
1113
struct FunctionRef {
    #[serde(rename = "$ref")]
    ref_path: String,
}

1114
#[derive(Debug, Serialize, Deserialize, PartialEq)]
drbh's avatar
drbh committed
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
struct Properties {
    #[serde(serialize_with = "serialize_function")]
    function: Vec<FunctionRef>,
}

fn serialize_function<S>(functions: &Vec<FunctionRef>, serializer: S) -> Result<S::Ok, S::Error>
where
    S: serde::Serializer,
{
    use serde::ser::SerializeStruct;
    let mut state = serializer.serialize_struct("Function", 1)?;
    state.serialize_field("anyOf", functions)?;
    state.end()
}

Nicolas Patry's avatar
Nicolas Patry committed
1130
#[derive(Clone, Debug, Deserialize, Serialize, ToSchema, Default, PartialEq)]
drbh's avatar
drbh committed
1131
1132
1133
1134
pub(crate) struct FunctionDefinition {
    #[serde(default)]
    pub description: Option<String>,
    pub name: String,
1135
1136
    #[serde(alias = "parameters")]
    pub arguments: serde_json::Value,
drbh's avatar
drbh committed
1137
1138
1139
}

#[derive(Clone, Debug, Deserialize, Serialize, ToSchema)]
Nicolas Patry's avatar
Nicolas Patry committed
1140
#[cfg_attr(test, derive(PartialEq))]
drbh's avatar
drbh committed
1141
1142
1143
1144
1145
1146
pub(crate) struct Tool {
    // The type of the tool. Currently, only 'function' is supported.
    #[schema(example = "function")]
    pub r#type: String,
    // Grab the tool as generic JSON for debugging purposes.
    pub function: FunctionDefinition,
1147
1148
}

1149
#[derive(Clone, Serialize, Deserialize, Default)]
1150
pub(crate) struct ChatTemplateInputs<'a> {
Nicolas Patry's avatar
Nicolas Patry committed
1151
    messages: Vec<TextMessage>,
1152
1153
    bos_token: Option<&'a str>,
    eos_token: Option<&'a str>,
1154
    add_generation_prompt: bool,
drbh's avatar
drbh committed
1155
    tools: Option<Vec<Tool>>,
1156
1157
}

Nicolas Patry's avatar
Nicolas Patry committed
1158
#[derive(Clone, Deserialize, Serialize, ToSchema, Default, Debug, PartialEq)]
drbh's avatar
drbh committed
1159
pub(crate) struct ToolCall {
1160
    pub id: String,
drbh's avatar
drbh committed
1161
1162
1163
1164
    pub r#type: String,
    pub function: FunctionDefinition,
}

Nicolas Patry's avatar
Nicolas Patry committed
1165
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
1166
pub struct Url {
Nicolas Patry's avatar
Nicolas Patry committed
1167
    url: String,
drbh's avatar
drbh committed
1168
1169
}

Nicolas Patry's avatar
Nicolas Patry committed
1170
1171
1172
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
#[serde(tag = "type")]
#[serde(rename_all = "snake_case")]
1173
1174
1175
pub enum MessageChunk {
    Text { text: String },
    ImageUrl { image_url: Url },
Nicolas Patry's avatar
Nicolas Patry committed
1176
1177
1178
1179
1180
1181
1182
}

#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
pub struct Message {
    #[schema(example = "user")]
    role: String,
    #[schema(example = "My name is David and I")]
1183
    pub content: MessageContent,
drbh's avatar
drbh committed
1184
    #[serde(default, skip_serializing_if = "Option::is_none")]
Nicolas Patry's avatar
Nicolas Patry committed
1185
1186
    #[schema(example = "\"David\"")]
    name: Option<String>,
drbh's avatar
drbh committed
1187
1188
}

1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
#[derive(Clone, Deserialize, Serialize, ToSchema, Debug, PartialEq)]
#[serde(untagged)]
pub enum MessageContent {
    SingleText(String),
    MultipleChunks(Vec<MessageChunk>),
}

// Pushing a chunk to a single text message will convert it to a multiple chunks message
impl MessageContent {
    pub fn push(&mut self, chunk: MessageChunk) {
        match self {
            MessageContent::SingleText(text) => {
drbh's avatar
drbh committed
1201
1202
1203
1204
                *self = MessageContent::MultipleChunks(vec![
                    MessageChunk::Text { text: text.clone() },
                    chunk,
                ]);
Nicolas Patry's avatar
Nicolas Patry committed
1205
            }
1206
1207
1208
1209
            MessageContent::MultipleChunks(chunks) => {
                chunks.push(chunk);
            }
        }
drbh's avatar
drbh committed
1210
1211
1212
    }
}

Nicolas Patry's avatar
Nicolas Patry committed
1213
1214
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
pub struct TextMessage {
1215
1216
1217
    #[schema(example = "user")]
    pub role: String,
    #[schema(example = "My name is David and I")]
Nicolas Patry's avatar
Nicolas Patry committed
1218
1219
1220
1221
1222
1223
1224
    pub content: String,
}

impl From<Message> for TextMessage {
    fn from(value: Message) -> Self {
        TextMessage {
            role: value.role,
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
            content: match value.content {
                MessageContent::SingleText(text) => text,
                MessageContent::MultipleChunks(chunks) => chunks
                    .into_iter()
                    .map(|chunk| match chunk {
                        MessageChunk::Text { text } => text,
                        MessageChunk::ImageUrl { image_url } => format!("![]({})", image_url.url),
                    })
                    .collect::<Vec<_>>()
                    .join(""),
            },
Nicolas Patry's avatar
Nicolas Patry committed
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
        }
    }
}

#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
pub struct ToolCallMessage {
    #[schema(example = "assistant")]
    role: String,
    tool_calls: Vec<ToolCall>,
}

#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
#[serde(untagged)]
pub(crate) enum OutputMessage {
    ChatMessage(TextMessage),
    ToolCall(ToolCallMessage),
1252
1253
}

1254
#[derive(Clone, Debug, Deserialize, ToSchema)]
1255
#[cfg_attr(test, derive(PartialEq))]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
1256
pub(crate) struct GenerateRequest {
1257
    #[schema(example = "My name is Olivier and I")]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
1258
1259
1260
    pub inputs: String,
    #[serde(default = "default_parameters")]
    pub parameters: GenerateParameters,
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270

    /// This is used internally because some requests
    /// already contain the templated input therefore
    /// we shouldn't add the special tokens.
    #[serde(default = "default_true", skip)]
    pub add_special_tokens: bool,
}

fn default_true() -> bool {
    true
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
1271
1272
}

1273
1274
1275
1276
1277
1278
1279
#[derive(Clone, Debug, Deserialize, ToSchema)]
pub(crate) struct CompatGenerateRequest {
    #[schema(example = "My name is Olivier and I")]
    pub inputs: String,
    #[serde(default = "default_parameters")]
    pub parameters: GenerateParameters,
    #[serde(default)]
OlivierDehaene's avatar
OlivierDehaene committed
1280
    #[schema(default = "false")]
1281
1282
1283
1284
1285
1286
1287
    pub stream: bool,
}

impl From<CompatGenerateRequest> for GenerateRequest {
    fn from(req: CompatGenerateRequest) -> Self {
        Self {
            inputs: req.inputs,
1288
            add_special_tokens: true,
1289
1290
1291
1292
1293
            parameters: req.parameters,
        }
    }
}

1294
1295
1296
#[derive(Debug, Serialize, ToSchema)]
pub struct PrefillToken {
    #[schema(example = 0)]
Nicolas Patry's avatar
Nicolas Patry committed
1297
    pub id: u32,
1298
    #[schema(example = "test")]
Nicolas Patry's avatar
Nicolas Patry committed
1299
    pub text: String,
1300
    #[schema(nullable = true, example = - 0.34)]
Nicolas Patry's avatar
Nicolas Patry committed
1301
    pub logprob: f32,
1302
1303
}

1304
#[derive(Debug, Serialize, ToSchema, Clone)]
1305
1306
pub struct Token {
    #[schema(example = 0)]
Nicolas Patry's avatar
Nicolas Patry committed
1307
    pub id: u32,
1308
    #[schema(example = "test")]
Nicolas Patry's avatar
Nicolas Patry committed
1309
    pub text: String,
1310
    #[schema(nullable = true, example = - 0.34)]
Nicolas Patry's avatar
Nicolas Patry committed
1311
    pub logprob: f32,
1312
    #[schema(example = "false")]
Nicolas Patry's avatar
Nicolas Patry committed
1313
    pub special: bool,
1314
1315
}

1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
#[derive(Debug, Serialize, ToSchema)]
pub struct SimpleToken {
    #[schema(example = 0)]
    id: u32,
    #[schema(example = "test")]
    text: String,
    #[schema(example = 0)]
    start: usize,
    #[schema(example = 2)]
    stop: usize,
}

1328
#[derive(Debug, Serialize, ToSchema, Clone)]
1329
#[serde(rename_all(serialize = "snake_case"))]
1330
#[schema(example = "Length")]
Nicolas Patry's avatar
Nicolas Patry committed
1331
pub enum FinishReason {
1332
1333
1334
1335
1336
1337
1338
1339
    #[schema(rename = "length")]
    Length,
    #[serde(rename = "eos_token")]
    #[schema(rename = "eos_token")]
    EndOfSequenceToken,
    #[schema(rename = "stop_sequence")]
    StopSequence,
}
1340

1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
impl std::fmt::Display for FinishReason {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            FinishReason::Length => write!(f, "length"),
            FinishReason::EndOfSequenceToken => write!(f, "eos_token"),
            FinishReason::StopSequence => write!(f, "stop_sequence"),
        }
    }
}

1351
1352
1353
1354
1355
1356
1357
1358
1359
impl FinishReason {
    pub fn format(&self, use_stop: bool) -> String {
        match self {
            FinishReason::EndOfSequenceToken if use_stop => "stop".to_string(),
            _ => self.to_string(),
        }
    }
}

1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
#[derive(Serialize, ToSchema)]
pub(crate) struct BestOfSequence {
    #[schema(example = "test")]
    pub generated_text: String,
    #[schema(example = "length")]
    pub finish_reason: FinishReason,
    #[schema(example = 1)]
    pub generated_tokens: u32,
    #[schema(nullable = true, example = 42)]
    pub seed: Option<u64>,
    pub prefill: Vec<PrefillToken>,
    pub tokens: Vec<Token>,
Nicolas Patry's avatar
Nicolas Patry committed
1372
1373
    #[serde(skip_serializing_if = "Vec::is_empty")]
    pub top_tokens: Vec<Vec<Token>>,
1374
1375
}

1376
#[derive(Serialize, ToSchema)]
OlivierDehaene's avatar
OlivierDehaene committed
1377
pub(crate) struct Details {
1378
1379
1380
    #[schema(example = "length")]
    pub finish_reason: FinishReason,
    #[schema(example = 1)]
OlivierDehaene's avatar
OlivierDehaene committed
1381
    pub generated_tokens: u32,
1382
    #[schema(nullable = true, example = 42)]
1383
    pub seed: Option<u64>,
1384
1385
    pub prefill: Vec<PrefillToken>,
    pub tokens: Vec<Token>,
1386
1387
    #[serde(skip_serializing_if = "Option::is_none")]
    pub best_of_sequences: Option<Vec<BestOfSequence>>,
Nicolas Patry's avatar
Nicolas Patry committed
1388
1389
    #[serde(skip_serializing_if = "Vec::is_empty")]
    pub top_tokens: Vec<Vec<Token>>,
OlivierDehaene's avatar
OlivierDehaene committed
1390
1391
}

1392
#[derive(Serialize, ToSchema)]
1393
pub(crate) struct GenerateResponse {
1394
    #[schema(example = "test")]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
1395
    pub generated_text: String,
OlivierDehaene's avatar
OlivierDehaene committed
1396
1397
    #[serde(skip_serializing_if = "Option::is_none")]
    pub details: Option<Details>,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
1398
}
1399

1400
1401
1402
1403
1404
1405
#[derive(Serialize, ToSchema)]
pub(crate) struct ChatTokenizeResponse {
    pub(crate) tokenize_response: TokenizeResponse,
    pub(crate) templated_text: String,
}

1406
1407
1408
1409
#[derive(Serialize, ToSchema)]
#[serde(transparent)]
pub(crate) struct TokenizeResponse(Vec<SimpleToken>);

1410
1411
1412
1413
1414
1415
#[derive(Serialize, ToSchema)]
pub(crate) struct StreamDetails {
    #[schema(example = "length")]
    pub finish_reason: FinishReason,
    #[schema(example = 1)]
    pub generated_tokens: u32,
1416
    #[schema(nullable = true, example = 42)]
1417
    pub seed: Option<u64>,
1418
1419
    #[schema(example = 1)]
    pub input_length: u32,
1420
1421
1422
}

#[derive(Serialize, ToSchema)]
1423
pub(crate) struct StreamResponse {
1424
    pub index: u32,
1425
    pub token: Token,
Nicolas Patry's avatar
Nicolas Patry committed
1426
1427
    #[serde(skip_serializing_if = "Vec::is_empty")]
    pub top_tokens: Vec<Token>,
1428
    #[schema(nullable = true, default = "null", example = "test")]
1429
    pub generated_text: Option<String>,
1430
1431
    #[schema(nullable = true, default = "null")]
    pub details: Option<StreamDetails>,
1432
1433
}

1434
#[derive(Serialize, ToSchema)]
1435
1436
pub(crate) struct ErrorResponse {
    pub error: String,
1437
    pub error_type: String,
1438
}
1439

drbh's avatar
drbh committed
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
#[derive(Serialize, Deserialize, ToSchema)]
pub(crate) struct ModelInfo {
    #[schema(example = "gpt2")]
    pub id: String,
    #[schema(example = "model")]
    pub object: String,
    #[schema(example = 1686935002)]
    pub created: u64,
    #[schema(example = "openai")]
    pub owned_by: String,
}

#[derive(Serialize, Deserialize, ToSchema)]
pub(crate) struct ModelsInfo {
    #[schema(example = "list")]
    pub object: String,
    pub data: Vec<ModelInfo>,
}

impl Default for ModelsInfo {
    fn default() -> Self {
        ModelsInfo {
            object: "list".to_string(),
            data: Vec::new(),
        }
    }
}

1468
#[cfg(test)]
1469
mod tests {
1470
    use super::*;
Nicolas Patry's avatar
Nicolas Patry committed
1471
    use serde_json::json;
1472

1473
    pub(crate) fn get_tokenizer() -> Tokenizer {
1474
1475
1476
        let api = hf_hub::api::sync::Api::new().unwrap();
        let repo = api.model("gpt2".to_string());
        let filename = repo.get("tokenizer.json").unwrap();
1477
        Tokenizer::Rust(tokenizers::Tokenizer::from_file(filename).unwrap())
1478
    }
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492

    #[test]
    fn test_hub_nested_tokens_tokenizer_config() {
        // this is a subset of the tokenizer.json file
        // in this case we expect the tokens to be encoded as simple strings
        let json_content = r#"{
            "chat_template": "test",
            "bos_token": "<|begin▁of▁sentence|>",
            "eos_token": "<|end▁of▁sentence|>"
        }"#;

        let config: HubTokenizerConfig = serde_json::from_str(json_content).unwrap();

        // check that we successfully parsed the tokens
1493
1494
1495
1496
        assert_eq!(
            config.chat_template,
            Some(ChatTemplateVersions::Single("test".to_string()))
        );
1497
1498
        assert_eq!(
            config.bos_token,
1499
1500
1501
1502
1503
1504
1505
1506
1507
            Some(TokenizerConfigToken::String(
                "<|begin▁of▁sentence|>".to_string()
            ))
        );
        assert_eq!(
            config.eos_token,
            Some(TokenizerConfigToken::String(
                "<|end▁of▁sentence|>".to_string()
            ))
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
        );

        // in this case we expect the tokens to be encoded as structured tokens
        // we want the content of the structured token
        let json_content = r#"{
            "chat_template": "test",
            "bos_token": {
              "__type": "AddedToken",
              "content": "<|begin▁of▁sentence|>",
              "lstrip": false,
              "normalized": true,
              "rstrip": false,
              "single_word": false
            },
            "eos_token": {
              "__type": "AddedToken",
              "content": "<|end▁of▁sentence|>",
              "lstrip": false,
              "normalized": true,
              "rstrip": false,
              "single_word": false
            }
        }"#;

        let config: HubTokenizerConfig = serde_json::from_str(json_content).unwrap();

        // check that we successfully parsed the tokens
1535
1536
1537
1538
        assert_eq!(
            config.chat_template,
            Some(ChatTemplateVersions::Single("test".to_string()))
        );
1539
1540
        assert_eq!(
            config.bos_token,
1541
1542
1543
1544
1545
1546
1547
1548
1549
            Some(TokenizerConfigToken::Object {
                content: "<|begin▁of▁sentence|>".to_string()
            })
        );
        assert_eq!(
            config.eos_token,
            Some(TokenizerConfigToken::Object {
                content: "<|end▁of▁sentence|>".to_string()
            })
1550
1551
        );
    }
Nicolas Patry's avatar
Nicolas Patry committed
1552
1553
1554

    #[test]
    fn test_chat_simple_string() {
Nicolas Patry's avatar
Nicolas Patry committed
1555
        let json = json!({
Nicolas Patry's avatar
Nicolas Patry committed
1556
            "model": "",
Nicolas Patry's avatar
Nicolas Patry committed
1557
1558
            "messages": [{
                "role": "user",
Nicolas Patry's avatar
Nicolas Patry committed
1559
                "content": "What is Deep Learning?"
Nicolas Patry's avatar
Nicolas Patry committed
1560
            }]
Nicolas Patry's avatar
Nicolas Patry committed
1561
1562
1563
1564
1565
1566
1567
        });
        let request: ChatRequest = serde_json::from_str(json.to_string().as_str()).unwrap();

        assert_eq!(
            request.messages[0],
            Message {
                role: "user".to_string(),
1568
                content: MessageContent::SingleText("What is Deep Learning?".to_string()),
Nicolas Patry's avatar
Nicolas Patry committed
1569
1570
1571
1572
1573
                name: None
            }
        );
    }

drbh's avatar
drbh committed
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
    #[test]
    fn test_message_content_append() {
        let mut content = MessageContent::SingleText("Initial text".to_string());
        let chunk = MessageChunk::Text {
            text: "Additional text".to_string(),
        };

        content.push(chunk);

        match content {
            MessageContent::MultipleChunks(chunks) => {
                assert_eq!(chunks.len(), 2);
                assert_eq!(
                    chunks[0],
                    MessageChunk::Text {
                        text: "Initial text".to_string()
                    }
                );
                assert_eq!(
                    chunks[1],
                    MessageChunk::Text {
                        text: "Additional text".to_string()
                    }
                );
            }
            _ => panic!("Expected MultipleChunks, but got a different variant"),
        }
    }

Nicolas Patry's avatar
Nicolas Patry committed
1603
1604
    #[test]
    fn test_chat_request() {
Nicolas Patry's avatar
Nicolas Patry committed
1605
        let json = json!({
Nicolas Patry's avatar
Nicolas Patry committed
1606
            "model": "",
Nicolas Patry's avatar
Nicolas Patry committed
1607
1608
            "messages": [{
                "role": "user",
Nicolas Patry's avatar
Nicolas Patry committed
1609
1610
                "content": [
                    {"type": "text", "text": "Whats in this image?"},
Nicolas Patry's avatar
Nicolas Patry committed
1611
                    {"type": "image_url", "image_url": {"url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png"}},
Nicolas Patry's avatar
Nicolas Patry committed
1612
                ]
Nicolas Patry's avatar
Nicolas Patry committed
1613
            }]
Nicolas Patry's avatar
Nicolas Patry committed
1614
1615
1616
1617
1618
1619
1620
        });
        let request: ChatRequest = serde_json::from_str(json.to_string().as_str()).unwrap();

        assert_eq!(
            request.messages[0],
            Message{
                role: "user".to_string(),
1621
1622
1623
1624
                content: MessageContent::MultipleChunks(vec![
                    MessageChunk::Text { text: "Whats in this image?".to_string() },
                    MessageChunk::ImageUrl { image_url: Url { url: "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png".to_string() }},
                ]),
Nicolas Patry's avatar
Nicolas Patry committed
1625
1626
1627
1628
                name: None
            }
        );
    }
Nicolas Patry's avatar
Nicolas Patry committed
1629
1630
1631
1632
1633

    #[test]
    fn text_message_convert() {
        let message = Message{
                role: "user".to_string(),
1634
1635
1636
1637
                content: MessageContent::MultipleChunks(vec![
                    MessageChunk::Text { text: "Whats in this image?".to_string() },
                    MessageChunk::ImageUrl { image_url: Url { url: "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png".to_string() } }
                ]),
Nicolas Patry's avatar
Nicolas Patry committed
1638
1639
1640
1641
1642
                name: None
            };
        let textmsg: TextMessage = message.into();
        assert_eq!(textmsg.content, "Whats in this image?![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png)");
    }
Nicolas Patry's avatar
Nicolas Patry committed
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663

    #[test]
    fn test_chat_stream_options() {
        let json = json!({
            "model": "",
            "stream_options": {"include_usage": true},
            "messages": [{
                "role": "user",
                "content": "Hello"
            }]
        });
        let request: ChatRequest = serde_json::from_str(json.to_string().as_str()).unwrap();

        assert!(matches!(
            request.stream_options,
            Some(StreamOptions {
                include_usage: true
            })
        ));
    }

Nicolas Patry's avatar
Nicolas Patry committed
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
    #[test]
    fn openai_output() {
        let message = OutputMessage::ChatMessage(TextMessage {
            role: "assistant".to_string(),
            content: "This is the answer".to_string(),
        });
        let serialized = serde_json::to_string(&message).unwrap();
        assert_eq!(
            serialized,
            r#"{"role":"assistant","content":"This is the answer"}"#
        );

        let message = OutputMessage::ToolCall(ToolCallMessage {
            role: "assistant".to_string(),
            tool_calls: vec![ToolCall {
                id: "0".to_string(),
                r#type: "function".to_string(),
                function: FunctionDefinition {
                    description: None,
                    name: "myfn".to_string(),
                    arguments: json!({
                        "format": "csv"
                    }),
                },
            }],
        });
        let serialized = serde_json::to_string(&message).unwrap();
        assert_eq!(
            serialized,
            r#"{"role":"assistant","tool_calls":[{"id":"0","type":"function","function":{"description":null,"name":"myfn","arguments":{"format":"csv"}}}]}"#
        );
    }
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732

    #[test]
    fn tool_choice_formats() {
        #[derive(Deserialize)]
        struct TestRequest {
            tool_choice: ToolChoice,
        }

        let de_none: TestRequest = serde_json::from_str(r#"{"tool_choice":"none"}"#).unwrap();
        assert_eq!(de_none.tool_choice, ToolChoice::NoTool);

        let de_auto: TestRequest = serde_json::from_str(r#"{"tool_choice":"auto"}"#).unwrap();
        assert_eq!(de_auto.tool_choice, ToolChoice::Auto);

        let de_required: TestRequest =
            serde_json::from_str(r#"{"tool_choice":"required"}"#).unwrap();
        assert_eq!(de_required.tool_choice, ToolChoice::Required);

        let de_named: TestRequest = serde_json::from_str(r#"{"tool_choice":"myfn"}"#).unwrap();
        assert_eq!(
            de_named.tool_choice,
            ToolChoice::Function(FunctionName {
                name: "myfn".to_string(),
            })
        );

        let de_openai_named: TestRequest = serde_json::from_str(
            r#"{"tool_choice":{"type":"function","function":{"name":"myfn"}}}"#,
        )
        .unwrap();
        assert_eq!(
            de_openai_named.tool_choice,
            ToolChoice::Function(FunctionName {
                name: "myfn".to_string(),
            })
        );
    }
1733
}