lib.rs 49.7 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
pub mod usage_stats;
Nicolas Patry's avatar
Nicolas Patry committed
12
mod vertex;
13

Nicolas Patry's avatar
Nicolas Patry committed
14
15
use crate::infer::{Infer, InferError};
use crate::server::prepare_chat_input;
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
16
use serde::{Deserialize, Serialize};
Nicolas Patry's avatar
Nicolas Patry committed
17
use tracing::warn;
18
use utoipa::ToSchema;
Olivier Dehaene's avatar
Olivier Dehaene committed
19
use validation::Validation;
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
20

21
22
23
24
25
26
27
#[derive(PartialEq)]
pub enum Attention {
    Paged,
    FlashDecoding,
    FlashInfer,
}

28
29
30
31
32
33
34
35
36
37
impl Attention {
    pub fn block_size(&self) -> u32 {
        match self {
            Attention::FlashDecoding => 256,
            Attention::FlashInfer => 1,
            Attention::Paged => 16,
        }
    }
}

38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
#[derive(Debug)]
pub struct ParseError;

impl std::fmt::Display for ParseError {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(f, "Cannot parse attention value")
    }
}
impl std::error::Error for ParseError {}

impl std::str::FromStr for Attention {
    type Err = ParseError;
    fn from_str(s: &str) -> Result<Self, Self::Err> {
        match s {
            "paged" => Ok(Attention::Paged),
            "flashdecoding" => Ok(Attention::FlashDecoding),
            "flashinfer" => Ok(Attention::FlashInfer),
            _ => Err(ParseError),
        }
    }
}

60
61
/// Hub type
#[derive(Clone, Debug, Deserialize)]
62
pub struct HubModelInfo {
63
64
65
66
67
68
    #[serde(rename(deserialize = "id"))]
    pub model_id: String,
    pub sha: Option<String>,
    pub pipeline_tag: Option<String>,
}

69
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
70
71
72
73
74
pub struct ChatTemplate {
    name: String,
    template: String,
}

75
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
76
77
78
79
80
81
#[serde(untagged)]
pub enum ChatTemplateVersions {
    Single(String),
    Multiple(Vec<ChatTemplate>),
}

82
83
use std::path::Path;

84
#[derive(Debug, Clone, Serialize, Deserialize, Default)]
85
pub struct HubTokenizerConfig {
86
    pub chat_template: Option<ChatTemplateVersions>,
87
    pub completion_template: Option<String>,
88
89
    pub bos_token: Option<TokenizerConfigToken>,
    pub eos_token: Option<TokenizerConfigToken>,
90
91
92
    pub tokenizer_class: Option<String>,
    pub add_bos_token: Option<bool>,
    pub add_eos_token: Option<bool>,
93
94
95
}

impl HubTokenizerConfig {
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
    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,
        }
116
117
118
    }
}

119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
#[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
138
139
140
141
142
143
144
145
#[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 {
146
147
148
149
    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
150
151
152
    }
}

153
#[derive(Clone, Debug, Deserialize, ToSchema, Serialize)]
Nicolas Patry's avatar
Nicolas Patry committed
154
#[cfg_attr(test, derive(PartialEq))]
drbh's avatar
drbh committed
155
156
#[serde(tag = "type", content = "value")]
pub(crate) enum GrammarType {
157
158
159
160
161
    /// 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
162
    #[serde(alias = "json_object")]
163
164
    #[schema(example = json ! ({"properties": {"location":{"type": "string"}}}))]
    Json(serde_json::Value),
drbh's avatar
drbh committed
165
166
167
168
    #[serde(rename = "regex")]
    Regex(String),
}

169
170
#[derive(Clone, Debug, Serialize, ToSchema)]
pub struct Info {
171
    /// Model info
172
173
174
175
    #[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
176
177
178
179
    // #[schema(example = "torch.float16")]
    // pub model_dtype: String,
    // #[schema(example = "cuda")]
    // pub model_device_type: String,
180
181
    #[schema(nullable = true, example = "text-generation")]
    pub model_pipeline_tag: Option<String>,
Nicolas Patry's avatar
Nicolas Patry committed
182

183
184
185
186
187
188
189
190
    /// 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
191
    pub max_input_tokens: usize,
192
193
194
195
    #[schema(example = "2048")]
    pub max_total_tokens: usize,
    #[schema(example = "2")]
    pub validation_workers: usize,
196
197
    #[schema(example = "32")]
    pub max_client_batch_size: usize,
Nicolas Patry's avatar
Nicolas Patry committed
198

199
    /// Router Info
200
201
    #[schema(example = "text-generation-router")]
    pub router: &'static str,
202
203
204
205
    #[schema(example = "0.5.0")]
    pub version: &'static str,
    #[schema(nullable = true, example = "null")]
    pub sha: Option<&'static str>,
206
207
    #[schema(nullable = true, example = "null")]
    pub docker_label: Option<&'static str>,
208
209
}

drbh's avatar
drbh committed
210
#[derive(Clone, Debug, Deserialize, ToSchema, Default)]
Nicolas Patry's avatar
Nicolas Patry committed
211
#[cfg_attr(test, derive(PartialEq))]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
212
pub(crate) struct GenerateParameters {
213
    /// Generate best_of sequences and return the one if the highest token logprobs.
214
215
216
    #[serde(default)]
    #[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 1)]
    pub best_of: Option<usize>,
217
218

    /// The value used to module the logits distribution.
219
220
221
222
223
224
225
226
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        nullable = true,
        default = "null",
        example = 0.5
    )]
    pub temperature: Option<f32>,
227
228
229

    /// The parameter for repetition penalty. 1.0 means no penalty.
    /// See [this paper](https://arxiv.org/pdf/1909.05858.pdf) for more details.
230
231
232
233
234
235
236
237
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        nullable = true,
        default = "null",
        example = 1.03
    )]
    pub repetition_penalty: Option<f32>,
238
239
240
241

    /// 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.
242
    #[serde(default)]
243
244
245
246
247
248
249
    #[schema(
        exclusive_minimum = -2.0,
        nullable = true,
        default = "null",
        example = 0.1
    )]
    pub frequency_penalty: Option<f32>,
250
251

    /// The number of highest probability vocabulary tokens to keep for top-k-filtering.
252
    #[serde(default)]
253
254
    #[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 10)]
    pub top_k: Option<i32>,
255
256

    /// Top-p value for nucleus sampling.
257
258
259
260
261
262
263
264
265
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        maximum = 1.0,
        nullable = true,
        default = "null",
        example = 0.95
    )]
    pub top_p: Option<f32>,
266
267
268

    /// Typical Decoding mass
    /// See [Typical Decoding for Natural Language Generation](https://arxiv.org/abs/2202.00666) for more information.
269
    #[serde(default)]
270
271
272
273
274
275
276
277
    #[schema(
        exclusive_minimum = 0.0,
        maximum = 1.0,
        nullable = true,
        default = "null",
        example = 0.95
    )]
    pub typical_p: Option<f32>,
278
279

    /// Activate logits sampling.
280
    #[serde(default)]
281
    #[schema(default = "false", example = true)]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
282
    pub do_sample: bool,
283
284

    /// Maximum number of tokens to generate.
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
285
    #[serde(default = "default_max_new_tokens")]
286
    #[schema(nullable = true, default = "100", example = "20")]
287
    pub max_new_tokens: Option<u32>,
288
289

    /// Whether to prepend the prompt to the generated text
OlivierDehaene's avatar
OlivierDehaene committed
290
    #[serde(default)]
291
    #[schema(nullable = true, default = "null", example = false)]
292
    pub return_full_text: Option<bool>,
293
294

    /// Stop generating tokens if a member of `stop` is generated.
295
    #[serde(default)]
296
    #[schema(inline, max_items = 4, example = json ! (["photographer"]))]
297
    pub stop: Vec<String>,
298
299

    /// Truncate inputs tokens to the given size.
OlivierDehaene's avatar
OlivierDehaene committed
300
    #[serde(default)]
301
    #[schema(nullable = true, default = "null", example = "null")]
302
    pub truncate: Option<usize>,
303
304

    /// Watermarking with [A Watermark for Large Language Models](https://arxiv.org/abs/2301.10226).
305
    #[serde(default)]
306
307
    #[schema(default = "false", example = true)]
    pub watermark: bool,
308
309

    /// Whether to return generation details.
310
    #[serde(default)]
311
    #[schema(default = "true")]
OlivierDehaene's avatar
OlivierDehaene committed
312
    pub details: bool,
313
314

    /// Whether to return decoder input token logprobs and ids.
315
    #[serde(default)]
316
    #[schema(default = "false")]
317
    pub decoder_input_details: bool,
318
319

    /// Random sampling seed.
320
    #[serde(default)]
321
322
323
324
325
326
    #[schema(
        exclusive_minimum = 0,
        nullable = true,
        default = "null",
        example = "null"
    )]
327
    pub seed: Option<u64>,
328
329

    /// The number of highest probability vocabulary tokens to keep for top-n-filtering.
Nicolas Patry's avatar
Nicolas Patry committed
330
331
332
    #[serde(default)]
    #[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 5)]
    pub top_n_tokens: Option<u32>,
333
334

    /// Grammar constraints for the generation.
drbh's avatar
drbh committed
335
    #[serde(default)]
336
    #[schema(nullable = true, default = "null", example = "null")]
drbh's avatar
drbh committed
337
    pub grammar: Option<GrammarType>,
drbh's avatar
drbh committed
338
339
340
341
342

    /// 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
343
344
}

345
fn default_max_new_tokens() -> Option<u32> {
346
    Some(100)
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
347
348
349
350
}

fn default_parameters() -> GenerateParameters {
    GenerateParameters {
351
        best_of: None,
352
353
        temperature: None,
        repetition_penalty: None,
354
        frequency_penalty: None,
355
356
        top_k: None,
        top_p: None,
357
        typical_p: None,
358
        do_sample: true,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
359
        max_new_tokens: default_max_new_tokens(),
360
        return_full_text: None,
361
        stop: Vec::new(),
362
        truncate: None,
363
        watermark: false,
OlivierDehaene's avatar
OlivierDehaene committed
364
        details: false,
365
        decoder_input_details: false,
366
        seed: None,
Nicolas Patry's avatar
Nicolas Patry committed
367
        top_n_tokens: None,
drbh's avatar
drbh committed
368
        grammar: None,
drbh's avatar
drbh committed
369
        adapter_id: None,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
370
371
372
    }
}

373
374
375
376
377
378
379
380
381
382
383
384
385
#[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;
386

387
    fn try_from(value: PromptDeserializer) -> Result<Self, Self::Error> {
388
        match value {
389
390
391
392
393
394
395
396
397
398
399
            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))
                }
            }
400
401
402
403
        }
    }
}

404
405
406
407
408
#[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.
409
    pub model: Option<String>,
410
411
412

    /// The prompt to generate completions for.
    #[schema(example = "What is Deep Learning?")]
413
    pub prompt: Prompt,
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450

    /// The maximum number of tokens that can be generated in the chat completion.
    #[serde(default)]
    #[schema(default = "32")]
    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>,
451
452
453
454
455

    /// Up to 4 sequences where the API will stop generating further tokens.
    #[serde(default)]
    #[schema(nullable = true, example = "null")]
    pub stop: Option<Vec<String>>,
456
457
}

458
459
460
461
462
463
464
465
466
#[derive(Clone, Serialize, ToSchema)]
#[serde(tag = "object")]
enum Completion {
    #[serde(rename = "text_completion")]
    Chunk(Chunk),
    #[serde(rename = "text_completion")]
    Final(CompletionFinal),
}

467
#[derive(Clone, Deserialize, Serialize, ToSchema, Default)]
468
pub(crate) struct CompletionFinal {
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
    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,
}

487
488
489
490
491
492
493
494
495
#[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,
}

496
#[derive(Clone, Deserialize, Serialize, ToSchema)]
497
498
pub(crate) struct ChatCompletion {
    pub id: String,
499
    #[schema(example = "1706270835")]
500
    pub created: u64,
501
    #[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
502
503
504
505
506
507
    pub model: String,
    pub system_fingerprint: String,
    pub choices: Vec<ChatCompletionComplete>,
    pub usage: Usage,
}

508
#[derive(Clone, Deserialize, Serialize, ToSchema)]
509
510
pub(crate) struct ChatCompletionComplete {
    pub index: u32,
Nicolas Patry's avatar
Nicolas Patry committed
511
    pub message: OutputMessage,
512
    pub logprobs: Option<ChatCompletionLogprobs>,
513
514
515
    pub finish_reason: String,
}

516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
#[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;
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558

        // 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
559
560
561
562
563
564
                        .into_iter()
                        .map(|t| ChatCompletionTopLogprob {
                            token: t.text,
                            logprob: t.logprob,
                        })
                        .collect(),
565
566
567
568
569
570
                    None => vec![], // Handle the case where there are no top tokens
                },
            })
            .collect();

        Self { content }
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
    }
}

#[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,
}

587
#[derive(Clone, Deserialize, Serialize, ToSchema, Default)]
588
589
590
591
592
593
pub(crate) struct Usage {
    pub prompt_tokens: u32,
    pub completion_tokens: u32,
    pub total_tokens: u32,
}

594
595
596
597
598
599
600
601
602
#[derive(Clone, Serialize, ToSchema)]
#[serde(tag = "object")]
enum CompletionType {
    #[serde(rename = "chat.completion.chunk")]
    ChatCompletionChunk(ChatCompletionChunk),
    #[serde(rename = "chat.completion")]
    ChatCompletion(ChatCompletion),
}

603
604
605
606
impl ChatCompletion {
    pub(crate) fn new(
        model: String,
        system_fingerprint: String,
drbh's avatar
drbh committed
607
        output: Option<String>,
608
609
610
        created: u64,
        details: Details,
        return_logprobs: bool,
611
        tool_calls: Option<Vec<ToolCall>>,
612
    ) -> Self {
Nicolas Patry's avatar
Nicolas Patry committed
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
        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(),
                })
            }
        };
637
638
639
640
641
642
643
        Self {
            id: String::new(),
            created,
            model,
            system_fingerprint,
            choices: vec![ChatCompletionComplete {
                index: 0,
Nicolas Patry's avatar
Nicolas Patry committed
644
                message,
645
                logprobs: return_logprobs
646
                    .then(|| ChatCompletionLogprobs::from((details.tokens, details.top_tokens))),
647
                finish_reason: details.finish_reason.format(true),
648
649
650
651
652
653
654
655
656
            }],
            usage: Usage {
                prompt_tokens: details.prefill.len() as u32,
                completion_tokens: details.generated_tokens,
                total_tokens: details.prefill.len() as u32 + details.generated_tokens,
            },
        }
    }
}
657
#[derive(Clone, Serialize, ToSchema)]
658
659
pub(crate) struct ChatCompletionChunk {
    pub id: String,
660
    #[schema(example = "1706270978")]
661
    pub created: u64,
662
    #[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
663
664
665
    pub model: String,
    pub system_fingerprint: String,
    pub choices: Vec<ChatCompletionChoice>,
Nicolas Patry's avatar
Nicolas Patry committed
666
    pub usage: Option<Usage>,
667
668
}

669
#[derive(Clone, Serialize, ToSchema)]
670
671
672
pub(crate) struct ChatCompletionChoice {
    pub index: u32,
    pub delta: ChatCompletionDelta,
673
    pub logprobs: Option<ChatCompletionLogprobs>,
674
675
676
    pub finish_reason: Option<String>,
}

Nicolas Patry's avatar
Nicolas Patry committed
677
678
679
680
681
682
683
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
pub struct ToolCallDelta {
    #[schema(example = "assistant")]
    role: String,
    tool_calls: DeltaToolCall,
}

684
685
#[derive(Clone, Debug, Serialize, ToSchema)]
#[serde(untagged)]
Nicolas Patry's avatar
Nicolas Patry committed
686
687
688
enum ChatCompletionDelta {
    Chat(TextMessage),
    Tool(ToolCallDelta),
drbh's avatar
drbh committed
689
690
}

Nicolas Patry's avatar
Nicolas Patry committed
691
#[derive(Clone, Deserialize, Serialize, ToSchema, Debug, PartialEq)]
drbh's avatar
drbh committed
692
693
694
695
696
697
698
pub(crate) struct DeltaToolCall {
    pub index: u32,
    pub id: String,
    pub r#type: String,
    pub function: Function,
}

Nicolas Patry's avatar
Nicolas Patry committed
699
#[derive(Clone, Deserialize, Serialize, ToSchema, Debug, PartialEq)]
drbh's avatar
drbh committed
700
701
702
pub(crate) struct Function {
    pub name: Option<String>,
    pub arguments: String,
703
704
}

drbh's avatar
drbh committed
705
#[allow(clippy::too_many_arguments)]
706
707
708
709
impl ChatCompletionChunk {
    pub(crate) fn new(
        model: String,
        system_fingerprint: String,
drbh's avatar
drbh committed
710
711
        delta: Option<String>,
        tool_calls: Option<Vec<String>>,
712
        created: u64,
713
        logprobs: Option<ChatCompletionLogprobs>,
714
        finish_reason: Option<String>,
Nicolas Patry's avatar
Nicolas Patry committed
715
        usage: Option<Usage>,
716
    ) -> Self {
717
        let delta = match (delta, tool_calls) {
Nicolas Patry's avatar
Nicolas Patry committed
718
719
720
721
722
723
724
            (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 {
725
726
727
728
729
730
731
                    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
732
733
734
735
736
737
                },
            }),
            (None, None) => ChatCompletionDelta::Chat(TextMessage {
                role: "assistant".to_string(),
                content: "".to_string(),
            }),
738
        };
739
740
741
742
743
744
        Self {
            id: String::new(),
            created,
            model,
            system_fingerprint,
            choices: vec![ChatCompletionChoice {
745
                index: 0,
746
                delta,
747
748
749
                logprobs,
                finish_reason,
            }],
Nicolas Patry's avatar
Nicolas Patry committed
750
            usage,
751
752
753
754
755
        }
    }
}

#[derive(Clone, Deserialize, ToSchema, Serialize)]
Nicolas Patry's avatar
Nicolas Patry committed
756
#[cfg_attr(test, derive(Debug, PartialEq, Default))]
757
pub(crate) struct ChatRequest {
758
    #[schema(example = "mistralai/Mistral-7B-Instruct-v0.2")]
drbh's avatar
drbh committed
759
    /// [UNUSED] ID of the model to use. See the model endpoint compatibility table for details on which models work with the Chat API.
760
    pub model: Option<String>,
drbh's avatar
drbh committed
761

762
    /// A list of messages comprising the conversation so far.
drbh's avatar
drbh committed
763
    #[schema(example = "[{\"role\": \"user\", \"content\": \"What is Deep Learning?\"}]")]
764
765
766
767
768
    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)]
769
    #[schema(example = "1.0")]
770
771
772
773
774
775
776
777
778
779
780
781
782
783
    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)]
784
    #[schema(example = "false")]
785
786
787
788
789
    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)]
790
    #[schema(example = "5")]
791
792
793
794
    pub top_logprobs: Option<u32>,

    /// The maximum number of tokens that can be generated in the chat completion.
    #[serde(default)]
795
    #[schema(example = "32")]
796
797
798
799
800
801
    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)]
802
    #[schema(nullable = true, example = "2")]
803
804
805
806
807
    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)]
808
    #[schema(nullable = true, example = 0.1)]
809
810
    pub presence_penalty: Option<f32>,

811
812
813
814
815
    /// Up to 4 sequences where the API will stop generating further tokens.
    #[serde(default)]
    #[schema(nullable = true, example = "null")]
    pub stop: Option<Vec<String>>,

816
817
818
819
820
    #[serde(default = "bool::default")]
    pub stream: bool,

    #[schema(nullable = true, example = 42)]
    pub seed: Option<u64>,
821
822
823
824
825
826

    /// 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)]
827
    #[schema(nullable = true, example = 1.0)]
828
829
830
831
832
    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)]
833
    #[schema(nullable = true, example = 0.95)]
834
    pub top_p: Option<f32>,
drbh's avatar
drbh committed
835
836
837
838
839
840
841
842

    /// 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
843
    #[serde(default)]
drbh's avatar
drbh committed
844
845
    #[schema(
        nullable = true,
drbh's avatar
drbh committed
846
        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
847
848
849
850
851
852
    )]
    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)]
    #[schema(nullable = true, example = "null")]
drbh's avatar
drbh committed
853
    pub tool_choice: ToolChoice,
drbh's avatar
drbh committed
854
855
856
857
858
859
860

    /// 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>,
861
862
863
864
865

    /// A guideline to be used in the chat_template
    #[serde(default)]
    #[schema(nullable = true, default = "null", example = "null")]
    pub guideline: Option<String>,
Nicolas Patry's avatar
Nicolas Patry committed
866
867
868
869
870
871
872

    /// 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
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
impl ChatRequest {
    fn try_into_generate(self, infer: &Infer) -> Result<(GenerateRequest, bool), InferError> {
        let ChatRequest {
            model,
            max_tokens,
            messages,
            seed,
            stop,
            stream,
            tools,
            tool_choice,
            tool_prompt,
            temperature,
            response_format,
            guideline,
            presence_penalty,
            frequency_penalty,
            top_p,
            top_logprobs,
            ..
        } = self;

        let repetition_penalty = presence_penalty.map(|x| x + 2.0);
        let max_new_tokens = max_tokens.or(Some(100));
        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),
        };
        let (inputs, grammar, using_tools) = prepare_chat_input(
            infer,
            response_format,
            tools,
            tool_choice,
            &tool_prompt,
            guideline,
            messages,
        )?;

        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,
                    decoder_input_details: !stream,
                    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
947
#[derive(Clone, Deserialize, ToSchema, Serialize)]
Nicolas Patry's avatar
Nicolas Patry committed
948
#[cfg_attr(test, derive(Debug, PartialEq))]
Nicolas Patry's avatar
Nicolas Patry committed
949
950
951
952
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
953
954
}

drbh's avatar
drbh committed
955
956
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
957
}
958
959
960
961

#[derive(Clone, Debug, Deserialize, PartialEq, Serialize, ToSchema)]
#[serde(untagged)]
pub enum ToolType {
drbh's avatar
drbh committed
962
    OneOf,
963
964
    FunctionName(String),
    Function { function: FunctionName },
drbh's avatar
drbh committed
965
    NoTool,
drbh's avatar
drbh committed
966
967
}

968
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, ToSchema)]
969
970
971
972
pub struct FunctionName {
    pub name: String,
}

drbh's avatar
drbh committed
973
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, Default, ToSchema)]
974
975
#[serde(from = "ToolTypeDeserializer")]
pub struct ToolChoice(pub Option<ToolType>);
drbh's avatar
drbh committed
976

977
978
979
#[derive(Deserialize)]
#[serde(untagged)]
enum ToolTypeDeserializer {
drbh's avatar
drbh committed
980
981
    String(String),
    ToolType(ToolType),
982
}
drbh's avatar
drbh committed
983

984
985
impl From<ToolTypeDeserializer> for ToolChoice {
    fn from(value: ToolTypeDeserializer) -> Self {
drbh's avatar
drbh committed
986
        match value {
drbh's avatar
drbh committed
987
988
989
990
            ToolTypeDeserializer::String(s) => match s.as_str() {
                "none" => ToolChoice(Some(ToolType::NoTool)),
                "auto" => ToolChoice(Some(ToolType::OneOf)),
                _ => ToolChoice(Some(ToolType::FunctionName(s))),
drbh's avatar
drbh committed
991
            },
drbh's avatar
drbh committed
992
            ToolTypeDeserializer::ToolType(tool_type) => ToolChoice(Some(tool_type)),
drbh's avatar
drbh committed
993
994
995
996
        }
    }
}

997
#[derive(Debug, Deserialize, Serialize, ToSchema, PartialEq)]
drbh's avatar
drbh committed
998
pub struct JsonSchemaTool {
drbh's avatar
drbh committed
999
1000
1001
1002
1003
    #[serde(flatten)]
    functions_map: FunctionsMap,
    properties: Properties,
}

1004
#[derive(Debug, Serialize, Deserialize, PartialEq)]
drbh's avatar
drbh committed
1005
1006
1007
1008
1009
struct FunctionsMap {
    #[serde(rename = "$functions")]
    functions: std::collections::HashMap<String, serde_json::Value>,
}

1010
#[derive(Debug, Serialize, Deserialize, PartialEq)]
drbh's avatar
drbh committed
1011
1012
1013
1014
1015
struct FunctionRef {
    #[serde(rename = "$ref")]
    ref_path: String,
}

1016
#[derive(Debug, Serialize, Deserialize, PartialEq)]
drbh's avatar
drbh committed
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
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
1032
#[derive(Clone, Debug, Deserialize, Serialize, ToSchema, Default, PartialEq)]
drbh's avatar
drbh committed
1033
1034
1035
1036
pub(crate) struct FunctionDefinition {
    #[serde(default)]
    pub description: Option<String>,
    pub name: String,
1037
1038
    #[serde(alias = "parameters")]
    pub arguments: serde_json::Value,
drbh's avatar
drbh committed
1039
1040
1041
}

#[derive(Clone, Debug, Deserialize, Serialize, ToSchema)]
Nicolas Patry's avatar
Nicolas Patry committed
1042
#[cfg_attr(test, derive(PartialEq))]
drbh's avatar
drbh committed
1043
1044
1045
1046
1047
1048
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,
1049
1050
}

1051
#[derive(Clone, Serialize, Deserialize, Default)]
1052
pub(crate) struct ChatTemplateInputs<'a> {
Nicolas Patry's avatar
Nicolas Patry committed
1053
    messages: Vec<TextMessage>,
1054
1055
    bos_token: Option<&'a str>,
    eos_token: Option<&'a str>,
1056
    add_generation_prompt: bool,
drbh's avatar
drbh committed
1057
    tools: Option<Vec<Tool>>,
1058
    guideline: Option<&'a str>,
1059
1060
}

Nicolas Patry's avatar
Nicolas Patry committed
1061
#[derive(Clone, Deserialize, Serialize, ToSchema, Default, Debug, PartialEq)]
drbh's avatar
drbh committed
1062
pub(crate) struct ToolCall {
1063
    pub id: String,
drbh's avatar
drbh committed
1064
1065
1066
1067
    pub r#type: String,
    pub function: FunctionDefinition,
}

Nicolas Patry's avatar
Nicolas Patry committed
1068
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
1069
pub struct Url {
Nicolas Patry's avatar
Nicolas Patry committed
1070
    url: String,
drbh's avatar
drbh committed
1071
1072
}

Nicolas Patry's avatar
Nicolas Patry committed
1073
1074
1075
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
#[serde(tag = "type")]
#[serde(rename_all = "snake_case")]
1076
1077
1078
pub enum MessageChunk {
    Text { text: String },
    ImageUrl { image_url: Url },
Nicolas Patry's avatar
Nicolas Patry committed
1079
1080
1081
1082
1083
1084
1085
}

#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
pub struct Message {
    #[schema(example = "user")]
    role: String,
    #[schema(example = "My name is David and I")]
1086
    pub content: MessageContent,
drbh's avatar
drbh committed
1087
    #[serde(default, skip_serializing_if = "Option::is_none")]
Nicolas Patry's avatar
Nicolas Patry committed
1088
1089
    #[schema(example = "\"David\"")]
    name: Option<String>,
drbh's avatar
drbh committed
1090
1091
}

1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
#[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
1104
1105
1106
1107
                *self = MessageContent::MultipleChunks(vec![
                    MessageChunk::Text { text: text.clone() },
                    chunk,
                ]);
Nicolas Patry's avatar
Nicolas Patry committed
1108
            }
1109
1110
1111
1112
            MessageContent::MultipleChunks(chunks) => {
                chunks.push(chunk);
            }
        }
drbh's avatar
drbh committed
1113
1114
1115
    }
}

Nicolas Patry's avatar
Nicolas Patry committed
1116
1117
#[derive(Clone, Deserialize, ToSchema, Serialize, Debug, PartialEq)]
pub struct TextMessage {
1118
1119
1120
    #[schema(example = "user")]
    pub role: String,
    #[schema(example = "My name is David and I")]
Nicolas Patry's avatar
Nicolas Patry committed
1121
1122
1123
1124
1125
1126
1127
    pub content: String,
}

impl From<Message> for TextMessage {
    fn from(value: Message) -> Self {
        TextMessage {
            role: value.role,
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
            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
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
        }
    }
}

#[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),
1155
1156
}

1157
#[derive(Clone, Debug, Deserialize, ToSchema)]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
1158
pub(crate) struct GenerateRequest {
1159
    #[schema(example = "My name is Olivier and I")]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
1160
1161
1162
    pub inputs: String,
    #[serde(default = "default_parameters")]
    pub parameters: GenerateParameters,
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172

    /// 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
1173
1174
}

1175
1176
1177
1178
1179
1180
1181
#[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
1182
    #[schema(default = "false")]
1183
1184
1185
1186
1187
1188
1189
    pub stream: bool,
}

impl From<CompatGenerateRequest> for GenerateRequest {
    fn from(req: CompatGenerateRequest) -> Self {
        Self {
            inputs: req.inputs,
1190
            add_special_tokens: true,
1191
1192
1193
1194
1195
            parameters: req.parameters,
        }
    }
}

1196
1197
1198
#[derive(Debug, Serialize, ToSchema)]
pub struct PrefillToken {
    #[schema(example = 0)]
Nicolas Patry's avatar
Nicolas Patry committed
1199
    pub id: u32,
1200
    #[schema(example = "test")]
Nicolas Patry's avatar
Nicolas Patry committed
1201
    pub text: String,
1202
    #[schema(nullable = true, example = - 0.34)]
Nicolas Patry's avatar
Nicolas Patry committed
1203
    pub logprob: f32,
1204
1205
}

1206
#[derive(Debug, Serialize, ToSchema, Clone)]
1207
1208
pub struct Token {
    #[schema(example = 0)]
Nicolas Patry's avatar
Nicolas Patry committed
1209
    pub id: u32,
1210
    #[schema(example = "test")]
Nicolas Patry's avatar
Nicolas Patry committed
1211
    pub text: String,
1212
    #[schema(nullable = true, example = - 0.34)]
Nicolas Patry's avatar
Nicolas Patry committed
1213
    pub logprob: f32,
1214
    #[schema(example = "false")]
Nicolas Patry's avatar
Nicolas Patry committed
1215
    pub special: bool,
1216
1217
}

1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
#[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,
}

OlivierDehaene's avatar
OlivierDehaene committed
1230
#[derive(Debug, Serialize, ToSchema)]
1231
#[serde(rename_all(serialize = "snake_case"))]
1232
#[schema(example = "Length")]
Nicolas Patry's avatar
Nicolas Patry committed
1233
pub enum FinishReason {
1234
1235
1236
1237
1238
1239
1240
1241
    #[schema(rename = "length")]
    Length,
    #[serde(rename = "eos_token")]
    #[schema(rename = "eos_token")]
    EndOfSequenceToken,
    #[schema(rename = "stop_sequence")]
    StopSequence,
}
1242

1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
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"),
        }
    }
}

1253
1254
1255
1256
1257
1258
1259
1260
1261
impl FinishReason {
    pub fn format(&self, use_stop: bool) -> String {
        match self {
            FinishReason::EndOfSequenceToken if use_stop => "stop".to_string(),
            _ => self.to_string(),
        }
    }
}

1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
#[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
1274
1275
    #[serde(skip_serializing_if = "Vec::is_empty")]
    pub top_tokens: Vec<Vec<Token>>,
1276
1277
}

1278
#[derive(Serialize, ToSchema)]
OlivierDehaene's avatar
OlivierDehaene committed
1279
pub(crate) struct Details {
1280
1281
1282
    #[schema(example = "length")]
    pub finish_reason: FinishReason,
    #[schema(example = 1)]
OlivierDehaene's avatar
OlivierDehaene committed
1283
    pub generated_tokens: u32,
1284
    #[schema(nullable = true, example = 42)]
1285
    pub seed: Option<u64>,
1286
1287
    pub prefill: Vec<PrefillToken>,
    pub tokens: Vec<Token>,
1288
1289
    #[serde(skip_serializing_if = "Option::is_none")]
    pub best_of_sequences: Option<Vec<BestOfSequence>>,
Nicolas Patry's avatar
Nicolas Patry committed
1290
1291
    #[serde(skip_serializing_if = "Vec::is_empty")]
    pub top_tokens: Vec<Vec<Token>>,
OlivierDehaene's avatar
OlivierDehaene committed
1292
1293
}

1294
#[derive(Serialize, ToSchema)]
1295
pub(crate) struct GenerateResponse {
1296
    #[schema(example = "test")]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
1297
    pub generated_text: String,
OlivierDehaene's avatar
OlivierDehaene committed
1298
1299
    #[serde(skip_serializing_if = "Option::is_none")]
    pub details: Option<Details>,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
1300
}
1301

1302
1303
1304
1305
1306
1307
#[derive(Serialize, ToSchema)]
pub(crate) struct ChatTokenizeResponse {
    pub(crate) tokenize_response: TokenizeResponse,
    pub(crate) templated_text: String,
}

1308
1309
1310
1311
#[derive(Serialize, ToSchema)]
#[serde(transparent)]
pub(crate) struct TokenizeResponse(Vec<SimpleToken>);

1312
1313
1314
1315
1316
1317
#[derive(Serialize, ToSchema)]
pub(crate) struct StreamDetails {
    #[schema(example = "length")]
    pub finish_reason: FinishReason,
    #[schema(example = 1)]
    pub generated_tokens: u32,
1318
    #[schema(nullable = true, example = 42)]
1319
    pub seed: Option<u64>,
1320
1321
    #[schema(example = 1)]
    pub input_length: u32,
1322
1323
1324
}

#[derive(Serialize, ToSchema)]
1325
pub(crate) struct StreamResponse {
1326
    pub index: u32,
1327
    pub token: Token,
Nicolas Patry's avatar
Nicolas Patry committed
1328
1329
    #[serde(skip_serializing_if = "Vec::is_empty")]
    pub top_tokens: Vec<Token>,
1330
    #[schema(nullable = true, default = "null", example = "test")]
1331
    pub generated_text: Option<String>,
1332
1333
    #[schema(nullable = true, default = "null")]
    pub details: Option<StreamDetails>,
1334
1335
}

1336
#[derive(Serialize, ToSchema)]
1337
1338
pub(crate) struct ErrorResponse {
    pub error: String,
1339
    pub error_type: String,
1340
}
1341

drbh's avatar
drbh committed
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
#[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(),
        }
    }
}

1370
#[cfg(test)]
1371
mod tests {
1372
    use super::*;
Nicolas Patry's avatar
Nicolas Patry committed
1373
    use serde_json::json;
1374
1375
    use tokenizers::Tokenizer;

1376
    pub(crate) async fn get_tokenizer() -> Tokenizer {
1377
1378
1379
1380
        let api = hf_hub::api::sync::Api::new().unwrap();
        let repo = api.model("gpt2".to_string());
        let filename = repo.get("tokenizer.json").unwrap();
        Tokenizer::from_file(filename).unwrap()
1381
    }
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395

    #[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
1396
1397
1398
1399
        assert_eq!(
            config.chat_template,
            Some(ChatTemplateVersions::Single("test".to_string()))
        );
1400
1401
        assert_eq!(
            config.bos_token,
1402
1403
1404
1405
1406
1407
1408
1409
1410
            Some(TokenizerConfigToken::String(
                "<|begin▁of▁sentence|>".to_string()
            ))
        );
        assert_eq!(
            config.eos_token,
            Some(TokenizerConfigToken::String(
                "<|end▁of▁sentence|>".to_string()
            ))
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
        );

        // 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
1438
1439
1440
1441
        assert_eq!(
            config.chat_template,
            Some(ChatTemplateVersions::Single("test".to_string()))
        );
1442
1443
        assert_eq!(
            config.bos_token,
1444
1445
1446
1447
1448
1449
1450
1451
1452
            Some(TokenizerConfigToken::Object {
                content: "<|begin▁of▁sentence|>".to_string()
            })
        );
        assert_eq!(
            config.eos_token,
            Some(TokenizerConfigToken::Object {
                content: "<|end▁of▁sentence|>".to_string()
            })
1453
1454
        );
    }
Nicolas Patry's avatar
Nicolas Patry committed
1455
1456
1457

    #[test]
    fn test_chat_simple_string() {
Nicolas Patry's avatar
Nicolas Patry committed
1458
        let json = json!({
Nicolas Patry's avatar
Nicolas Patry committed
1459
            "model": "",
Nicolas Patry's avatar
Nicolas Patry committed
1460
1461
            "messages": [{
                "role": "user",
Nicolas Patry's avatar
Nicolas Patry committed
1462
                "content": "What is Deep Learning?"
Nicolas Patry's avatar
Nicolas Patry committed
1463
            }]
Nicolas Patry's avatar
Nicolas Patry committed
1464
1465
1466
1467
1468
1469
1470
        });
        let request: ChatRequest = serde_json::from_str(json.to_string().as_str()).unwrap();

        assert_eq!(
            request.messages[0],
            Message {
                role: "user".to_string(),
1471
                content: MessageContent::SingleText("What is Deep Learning?".to_string()),
Nicolas Patry's avatar
Nicolas Patry committed
1472
1473
1474
1475
1476
                name: None
            }
        );
    }

drbh's avatar
drbh committed
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
    #[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
1506
1507
    #[test]
    fn test_chat_request() {
Nicolas Patry's avatar
Nicolas Patry committed
1508
        let json = json!({
Nicolas Patry's avatar
Nicolas Patry committed
1509
            "model": "",
Nicolas Patry's avatar
Nicolas Patry committed
1510
1511
            "messages": [{
                "role": "user",
Nicolas Patry's avatar
Nicolas Patry committed
1512
1513
                "content": [
                    {"type": "text", "text": "Whats in this image?"},
Nicolas Patry's avatar
Nicolas Patry committed
1514
                    {"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
1515
                ]
Nicolas Patry's avatar
Nicolas Patry committed
1516
            }]
Nicolas Patry's avatar
Nicolas Patry committed
1517
1518
1519
1520
1521
1522
1523
        });
        let request: ChatRequest = serde_json::from_str(json.to_string().as_str()).unwrap();

        assert_eq!(
            request.messages[0],
            Message{
                role: "user".to_string(),
1524
1525
1526
1527
                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
1528
1529
1530
1531
                name: None
            }
        );
    }
Nicolas Patry's avatar
Nicolas Patry committed
1532
1533
1534
1535
1536

    #[test]
    fn text_message_convert() {
        let message = Message{
                role: "user".to_string(),
1537
1538
1539
1540
                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
1541
1542
1543
1544
1545
                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
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566

    #[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
1567
1568
1569
1570
1571
1572
1573
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
    #[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"}}}]}"#
        );
    }
1599
}