lib.rs 16.3 KB
Newer Older
1
mod health;
2
/// Text Generation Inference Webserver
3
mod infer;
4
mod queue;
Olivier Dehaene's avatar
Olivier Dehaene committed
5
pub mod server;
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
6
mod validation;
Olivier Dehaene's avatar
Olivier Dehaene committed
7

8
use infer::{Infer, InferError, InferStreamResponse};
9
use queue::{Entry, Queue};
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
10
use serde::{Deserialize, Serialize};
11
12
use tokio::sync::OwnedSemaphorePermit;
use tokio_stream::wrappers::UnboundedReceiverStream;
13
use utoipa::ToSchema;
Olivier Dehaene's avatar
Olivier Dehaene committed
14
use validation::Validation;
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
15

16
17
18
19
20
21
22
/// Type alias for generation responses
pub(crate) type GenerateStreamResponse = (
    OwnedSemaphorePermit,
    u32, // input_length
    UnboundedReceiverStream<Result<InferStreamResponse, InferError>>,
);

23
24
/// Hub type
#[derive(Clone, Debug, Deserialize)]
25
pub struct HubModelInfo {
26
27
28
29
30
31
    #[serde(rename(deserialize = "id"))]
    pub model_id: String,
    pub sha: Option<String>,
    pub pipeline_tag: Option<String>,
}

32
33
34
#[derive(Clone, Deserialize, Default)]
pub struct HubTokenizerConfig {
    pub chat_template: Option<String>,
35
36
    pub bos_token: Option<String>,
    pub eos_token: Option<String>,
37
38
39
40
41
42
43
44
45
}

impl HubTokenizerConfig {
    pub fn from_file(filename: &str) -> Self {
        let content = std::fs::read_to_string(filename).unwrap();
        serde_json::from_str(&content).unwrap_or_default()
    }
}

46
47
#[derive(Clone, Debug, Serialize, ToSchema)]
pub struct Info {
48
    /// Model info
49
50
51
52
    #[schema(example = "bigscience/blomm-560m")]
    pub model_id: String,
    #[schema(nullable = true, example = "e985a63cdc139290c5f700ff1929f0b5942cced2")]
    pub model_sha: Option<String>,
53
54
55
56
    #[schema(example = "torch.float16")]
    pub model_dtype: String,
    #[schema(example = "cuda")]
    pub model_device_type: String,
57
58
    #[schema(nullable = true, example = "text-generation")]
    pub model_pipeline_tag: Option<String>,
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
    /// 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")]
    pub max_input_length: usize,
    #[schema(example = "2048")]
    pub max_total_tokens: usize,
    #[schema(example = "1.2")]
    pub waiting_served_ratio: f32,
    #[schema(example = "32000")]
    pub max_batch_total_tokens: u32,
    #[schema(example = "20")]
    pub max_waiting_tokens: usize,
    #[schema(example = "2")]
    pub validation_workers: usize,
    /// Router Info
79
80
81
82
    #[schema(example = "0.5.0")]
    pub version: &'static str,
    #[schema(nullable = true, example = "null")]
    pub sha: Option<&'static str>,
83
84
    #[schema(nullable = true, example = "null")]
    pub docker_label: Option<&'static str>,
85
86
}

87
#[derive(Clone, Debug, Deserialize, ToSchema)]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
88
pub(crate) struct GenerateParameters {
89
90
91
    #[serde(default)]
    #[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 1)]
    pub best_of: Option<usize>,
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        nullable = true,
        default = "null",
        example = 0.5
    )]
    pub temperature: Option<f32>,
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        nullable = true,
        default = "null",
        example = 1.03
    )]
    pub repetition_penalty: Option<f32>,
    #[serde(default)]
    #[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 10)]
    pub top_k: Option<i32>,
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        maximum = 1.0,
        nullable = true,
        default = "null",
        example = 0.95
    )]
    pub top_p: Option<f32>,
120
    #[serde(default)]
121
122
123
124
125
126
127
128
129
    #[schema(
        exclusive_minimum = 0.0,
        maximum = 1.0,
        nullable = true,
        default = "null",
        example = 0.95
    )]
    pub typical_p: Option<f32>,
    #[serde(default)]
130
    #[schema(default = "false", example = true)]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
131
132
    pub do_sample: bool,
    #[serde(default = "default_max_new_tokens")]
133
    #[schema(nullable = true, default = "100", example = "20")]
134
    pub max_new_tokens: Option<u32>,
OlivierDehaene's avatar
OlivierDehaene committed
135
    #[serde(default)]
136
    #[schema(nullable = true, default = "null", example = false)]
137
138
    pub return_full_text: Option<bool>,
    #[serde(default)]
139
    #[schema(inline, max_items = 4, example = json ! (["photographer"]))]
140
    pub stop: Vec<String>,
OlivierDehaene's avatar
OlivierDehaene committed
141
    #[serde(default)]
142
    #[schema(nullable = true, default = "null", example = "null")]
143
144
    pub truncate: Option<usize>,
    #[serde(default)]
145
146
147
    #[schema(default = "false", example = true)]
    pub watermark: bool,
    #[serde(default)]
148
    #[schema(default = "true")]
OlivierDehaene's avatar
OlivierDehaene committed
149
    pub details: bool,
150
    #[serde(default)]
151
152
153
    #[schema(default = "true")]
    pub decoder_input_details: bool,
    #[serde(default)]
154
155
156
157
158
159
    #[schema(
        exclusive_minimum = 0,
        nullable = true,
        default = "null",
        example = "null"
    )]
160
    pub seed: Option<u64>,
Nicolas Patry's avatar
Nicolas Patry committed
161
162
163
    #[serde(default)]
    #[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 5)]
    pub top_n_tokens: Option<u32>,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
164
165
}

166
fn default_max_new_tokens() -> Option<u32> {
167
    Some(100)
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
168
169
170
171
}

fn default_parameters() -> GenerateParameters {
    GenerateParameters {
172
        best_of: None,
173
174
175
176
        temperature: None,
        repetition_penalty: None,
        top_k: None,
        top_p: None,
177
        typical_p: None,
178
        do_sample: true,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
179
        max_new_tokens: default_max_new_tokens(),
180
        return_full_text: None,
181
        stop: Vec::new(),
182
        truncate: None,
183
        watermark: false,
OlivierDehaene's avatar
OlivierDehaene committed
184
        details: false,
185
        decoder_input_details: false,
186
        seed: None,
Nicolas Patry's avatar
Nicolas Patry committed
187
        top_n_tokens: None,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
188
189
190
    }
}

191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
#[derive(Clone, Deserialize, Serialize)]
pub(crate) struct ChatCompletion {
    pub id: String,
    pub object: String,
    pub created: u64,
    pub model: String,
    pub system_fingerprint: String,
    pub choices: Vec<ChatCompletionComplete>,
    pub usage: Usage,
}

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

#[derive(Clone, Deserialize, Serialize)]
pub(crate) struct Usage {
    pub prompt_tokens: u32,
    pub completion_tokens: u32,
    pub total_tokens: u32,
}

impl ChatCompletion {
    pub(crate) fn new(
        model: String,
        system_fingerprint: String,
        output: String,
        created: u64,
        details: Details,
        return_logprobs: bool,
    ) -> Self {
        Self {
            id: String::new(),
            object: "text_completion".into(),
            created,
            model,
            system_fingerprint,
            choices: vec![ChatCompletionComplete {
                index: 0,
                message: Message {
                    role: "assistant".into(),
                    content: output,
                },
                logprobs: return_logprobs
                    .then(|| details.tokens.iter().map(|t| t.logprob).collect()),
                finish_reason: details.finish_reason.to_string(),
            }],
            usage: Usage {
                prompt_tokens: details.prefill.len() as u32,
                completion_tokens: details.generated_tokens,
                total_tokens: details.prefill.len() as u32 + details.generated_tokens,
            },
        }
    }
}

#[derive(Clone, Deserialize, Serialize)]
pub(crate) struct ChatCompletionChunk {
    pub id: String,
    pub object: String,
    pub created: u64,
    pub model: String,
    pub system_fingerprint: String,
    pub choices: Vec<ChatCompletionChoice>,
}

#[derive(Clone, Deserialize, Serialize)]
pub(crate) struct ChatCompletionChoice {
    pub index: u32,
    pub delta: ChatCompletionDelta,
    pub logprobs: Option<f32>,
    pub finish_reason: Option<String>,
}

#[derive(Clone, Debug, Deserialize, Serialize)]
pub(crate) struct ChatCompletionDelta {
    pub role: String,
    pub content: String,
}

impl ChatCompletionChunk {
    pub(crate) fn new(
        model: String,
        system_fingerprint: String,
        delta: String,
        created: u64,
        index: u32,
        logprobs: Option<f32>,
        finish_reason: Option<String>,
    ) -> Self {
        Self {
            id: String::new(),
            object: "text_completion".to_string(),
            created,
            model,
            system_fingerprint,
            choices: vec![ChatCompletionChoice {
                index,
                delta: ChatCompletionDelta {
                    role: "assistant".to_string(),
                    content: delta,
                },
                logprobs,
                finish_reason,
            }],
        }
    }
}

fn default_request_messages() -> Vec<Message> {
    vec![Message {
        role: "user".to_string(),
        content: "My name is David and I".to_string(),
    }]
}

#[derive(Clone, Deserialize, ToSchema, Serialize)]
pub(crate) struct ChatRequest {
    /// UNUSED
    #[schema(example = "bigscience/blomm-560m")]
    /// ID of the model to use. See the model endpoint compatibility table for details on which models work with the Chat API.
    pub model: String, /* NOTE: UNUSED */

    /// A list of messages comprising the conversation so far.
    #[serde(default = "default_request_messages")]
    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)]
    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)]
    pub logprobs: Option<bool>,

    /// UNUSED
    /// 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)]
    pub top_logprobs: Option<u32>,

    /// The maximum number of tokens that can be generated in the chat completion.
    #[serde(default)]
    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)]
    pub n: Option<u32>,

    /// UNUSED
    /// 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)]
    pub presence_penalty: Option<f32>,

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

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

370
371
372
373
374
375
376
#[derive(Clone, Serialize, Deserialize)]
pub(crate) struct ChatTemplateInputs<'a> {
    messages: Vec<Message>,
    bos_token: Option<&'a str>,
    eos_token: Option<&'a str>,
}

377
378
379
380
381
382
383
384
#[derive(Clone, Deserialize, ToSchema, Serialize)]
pub(crate) struct Message {
    #[schema(example = "user")]
    pub role: String,
    #[schema(example = "My name is David and I")]
    pub content: String,
}

385
#[derive(Clone, Debug, Deserialize, ToSchema)]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
386
pub(crate) struct GenerateRequest {
387
    #[schema(example = "My name is Olivier and I")]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
388
389
390
391
392
    pub inputs: String,
    #[serde(default = "default_parameters")]
    pub parameters: GenerateParameters,
}

393
394
395
396
397
398
399
#[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
400
    #[schema(default = "false")]
401
402
403
404
405
406
407
408
409
410
411
412
    pub stream: bool,
}

impl From<CompatGenerateRequest> for GenerateRequest {
    fn from(req: CompatGenerateRequest) -> Self {
        Self {
            inputs: req.inputs,
            parameters: req.parameters,
        }
    }
}

413
414
415
416
417
418
#[derive(Debug, Serialize, ToSchema)]
pub struct PrefillToken {
    #[schema(example = 0)]
    id: u32,
    #[schema(example = "test")]
    text: String,
419
    #[schema(nullable = true, example = - 0.34)]
420
421
422
    logprob: f32,
}

423
424
425
426
427
428
#[derive(Debug, Serialize, ToSchema)]
pub struct Token {
    #[schema(example = 0)]
    id: u32,
    #[schema(example = "test")]
    text: String,
429
    #[schema(nullable = true, example = - 0.34)]
430
    logprob: f32,
431
432
    #[schema(example = "false")]
    special: bool,
433
434
435
436
437
438
439
440
441
442
443
444
445
}

#[derive(Serialize, ToSchema)]
#[serde(rename_all(serialize = "snake_case"))]
pub(crate) enum FinishReason {
    #[schema(rename = "length")]
    Length,
    #[serde(rename = "eos_token")]
    #[schema(rename = "eos_token")]
    EndOfSequenceToken,
    #[schema(rename = "stop_sequence")]
    StopSequence,
}
446

447
448
449
450
451
452
453
454
455
456
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"),
        }
    }
}

457
458
459
460
461
462
463
464
465
466
467
468
#[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
469
470
    #[serde(skip_serializing_if = "Vec::is_empty")]
    pub top_tokens: Vec<Vec<Token>>,
471
472
}

473
#[derive(Serialize, ToSchema)]
OlivierDehaene's avatar
OlivierDehaene committed
474
pub(crate) struct Details {
475
476
477
    #[schema(example = "length")]
    pub finish_reason: FinishReason,
    #[schema(example = 1)]
OlivierDehaene's avatar
OlivierDehaene committed
478
    pub generated_tokens: u32,
479
    #[schema(nullable = true, example = 42)]
480
    pub seed: Option<u64>,
481
482
    pub prefill: Vec<PrefillToken>,
    pub tokens: Vec<Token>,
483
484
    #[serde(skip_serializing_if = "Option::is_none")]
    pub best_of_sequences: Option<Vec<BestOfSequence>>,
Nicolas Patry's avatar
Nicolas Patry committed
485
486
    #[serde(skip_serializing_if = "Vec::is_empty")]
    pub top_tokens: Vec<Vec<Token>>,
OlivierDehaene's avatar
OlivierDehaene committed
487
488
}

489
#[derive(Serialize, ToSchema)]
490
pub(crate) struct GenerateResponse {
491
    #[schema(example = "test")]
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
492
    pub generated_text: String,
OlivierDehaene's avatar
OlivierDehaene committed
493
494
    #[serde(skip_serializing_if = "Option::is_none")]
    pub details: Option<Details>,
Olivier Dehaene's avatar
v0.1.0  
Olivier Dehaene committed
495
}
496

497
498
499
500
501
502
#[derive(Serialize, ToSchema)]
pub(crate) struct StreamDetails {
    #[schema(example = "length")]
    pub finish_reason: FinishReason,
    #[schema(example = 1)]
    pub generated_tokens: u32,
503
    #[schema(nullable = true, example = 42)]
504
505
506
507
    pub seed: Option<u64>,
}

#[derive(Serialize, ToSchema)]
508
pub(crate) struct StreamResponse {
509
    pub index: u32,
510
    pub token: Token,
Nicolas Patry's avatar
Nicolas Patry committed
511
512
    #[serde(skip_serializing_if = "Vec::is_empty")]
    pub top_tokens: Vec<Token>,
513
    #[schema(nullable = true, default = "null", example = "test")]
514
    pub generated_text: Option<String>,
515
516
    #[schema(nullable = true, default = "null")]
    pub details: Option<StreamDetails>,
517
518
}

519
#[derive(Serialize, ToSchema)]
520
521
pub(crate) struct ErrorResponse {
    pub error: String,
522
    pub error_type: String,
523
}
524
525

#[cfg(test)]
526
mod tests {
527
528
529
    use std::io::Write;
    use tokenizers::Tokenizer;

530
    pub(crate) async fn get_tokenizer() -> Tokenizer {
531
532
        let filename = std::path::Path::new("tokenizer.json");
        if !filename.exists() {
533
534
535
536
537
538
            let content = reqwest::get("https://huggingface.co/gpt2/raw/main/tokenizer.json")
                .await
                .unwrap()
                .bytes()
                .await
                .unwrap();
539
540
            let tmp_filename = "tokenizer.json.temp";
            let mut file = std::fs::File::create(tmp_filename).unwrap();
541
            file.write_all(&content).unwrap();
542
543
544
545
            // Re-check if another process has written this file maybe.
            if !filename.exists() {
                std::fs::rename(tmp_filename, filename).unwrap()
            }
546
547
548
549
        }
        Tokenizer::from_file("tokenizer.json").unwrap()
    }
}