infer.rs 21.6 KB
Newer Older
1
2
/// Batching and inference logic
use crate::validation::{Validation, ValidationError};
3
use crate::{Entry, Queue, Token};
4
use crate::{GenerateRequest, PrefillToken};
5
use flume::r#async::RecvStream;
6
use flume::SendTimeoutError;
7
use futures::future::try_join_all;
8
use futures::stream::StreamExt;
9
use nohash_hasher::IntMap;
10
11
12
13
use std::sync::{
    atomic::{AtomicBool, Ordering},
    Arc,
};
14
use std::time::Duration;
15
use text_generation_client::{
16
    Batch, CachedBatch, ClientError, GeneratedText, Generation, PrefillTokens, ShardedClient,
17
18
};
use thiserror::Error;
19
use tokio::sync::{Notify, OwnedSemaphorePermit, Semaphore, TryAcquireError};
20
use tokio::time::Instant;
21
use tracing::{info_span, instrument, Instrument, Span};
22
23
24
25
26
27

/// Inference struct
#[derive(Clone)]
pub struct Infer {
    /// Validation
    validation: Validation,
28
29
    /// Request queue
    queue: Queue,
30
31
32
33
34
35
36
37
38
39
40
41
42
    /// Shared state
    shared: Arc<Shared>,
    /// Inference limit
    limit_concurrent_requests: Arc<Semaphore>,
}

/// Infer shared state
struct Shared {
    /// Batching background Tokio task notifier
    batching_task: Notify,
}

impl Infer {
43
    #[allow(clippy::too_many_arguments)]
44
45
46
    pub(crate) fn new(
        client: ShardedClient,
        validation: Validation,
47
48
        waiting_served_ratio: f32,
        max_batch_total_tokens: u32,
49
50
        max_waiting_tokens: usize,
        max_concurrent_requests: usize,
51
        requires_padding: bool,
52
        generation_health: Arc<AtomicBool>,
53
54
    ) -> Self {
        // Infer shared state
55
        let queue = Queue::new(requires_padding);
56
57
58
59
60
61
62
        let shared = Arc::new(Shared {
            batching_task: Notify::new(),
        });

        // Spawn batching background task that contains all the inference logic
        tokio::spawn(batching_task(
            client,
63
64
            waiting_served_ratio,
            max_batch_total_tokens,
65
            max_waiting_tokens,
66
            queue.clone(),
67
            shared.clone(),
68
            generation_health,
69
70
71
72
73
74
75
        ));

        // Inference limit with a semaphore
        let semaphore = Arc::new(Semaphore::new(max_concurrent_requests));

        Self {
            validation,
76
            queue,
77
78
79
80
81
            shared,
            limit_concurrent_requests: semaphore,
        }
    }

82
    /// Add a new request to the queue and return a stream of InferStreamResponse
83
    #[instrument(skip(self))]
84
85
86
    pub(crate) async fn generate_stream(
        &self,
        request: GenerateRequest,
87
88
89
90
91
92
93
    ) -> Result<
        (
            OwnedSemaphorePermit,
            RecvStream<Result<InferStreamResponse, InferError>>,
        ),
        InferError,
    > {
94
        // Limit concurrent requests by acquiring a permit from the semaphore
95
96
97
98
99
        let permit = self
            .clone()
            .limit_concurrent_requests
            .try_acquire_owned()
            .map_err(|err| {
100
                metrics::increment_counter!("tgi_request_failure", "err" => "overloaded");
101
102
103
                tracing::error!("{err}");
                err
            })?;
104
105

        // Validate request
106
107
108
109
110
        let valid_request = self.validation.validate(request).await.map_err(|err| {
            metrics::increment_counter!("tgi_request_failure", "err" => "validation");
            tracing::error!("{err}");
            err
        })?;
111
112

        // MPSC channel to communicate with the background batching task
113
        let (response_tx, response_rx) = flume::unbounded();
114

115
116
        // Append the request to the queue
        self.queue.append(Entry {
117
118
            request: valid_request,
            response_tx,
119
120
121
            span: Span::current(),
            temp_span: None,
            queue_time: Instant::now(),
122
123
124
            batch_time: None,
        });

125
        // Notify the background task that we have a new entry in the queue that needs
126
127
128
129
        // to be batched
        self.shared.batching_task.notify_one();

        // Return stream
130
        Ok((permit, response_rx.into_stream()))
131
132
    }

133
    /// Add a new request to the queue and return a InferResponse
134
    #[instrument(skip(self))]
135
136
137
138
    pub(crate) async fn generate(
        &self,
        request: GenerateRequest,
    ) -> Result<InferResponse, InferError> {
139
140
        // Create stream and keep semaphore permit as long as generate lives
        let (_permit, mut stream) = self.generate_stream(request).await?;
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160

        // Return values
        let mut result_prefill = Vec::new();
        let mut result_tokens = Vec::new();
        let mut result_generated_text = None;
        let mut result_start = None;
        let mut result_queued = None;

        // Iterate on stream
        while let Some(response) = stream.next().await {
            match response? {
                // Add prefill tokens
                InferStreamResponse::Prefill(tokens) => {
                    // Create Token objects
                    // We do that here instead of in the Python code as Rust for loops are faster
                    result_prefill = tokens
                        .ids
                        .into_iter()
                        .zip(tokens.logprobs.into_iter())
                        .zip(tokens.texts.into_iter())
161
                        .map(|((id, logprob), text)| PrefillToken { id, text, logprob })
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
                        .collect();
                }
                // Push last token
                InferStreamResponse::Token(token) => result_tokens.push(token),
                // Final message
                // Set return values
                InferStreamResponse::End {
                    token,
                    generated_text,
                    start,
                    queued,
                } => {
                    result_tokens.push(token);
                    result_generated_text = Some(generated_text);
                    result_start = Some(start);
                    result_queued = Some(queued)
                }
            }
        }

        // Check that we received a `InferStreamResponse::End` message
        if let (Some(generated_text), Some(queued), Some(start)) =
            (result_generated_text, result_queued, result_start)
        {
            Ok(InferResponse {
                prefill: result_prefill,
                tokens: result_tokens,
                generated_text,
                queued,
                start,
            })
        } else {
194
            let err = InferError::IncompleteGeneration;
195
            metrics::increment_counter!("tgi_request_failure", "err" => "incomplete");
196
197
            tracing::error!("{err}");
            Err(err)
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
    /// Add best_of new requests to the queue and return a InferResponse of the sequence with
    /// the highest log probability per token
    #[instrument(skip(self))]
    pub(crate) async fn generate_best_of(
        &self,
        request: GenerateRequest,
        best_of: usize,
    ) -> Result<(InferResponse, Vec<InferResponse>), InferError> {
        // validate  best_of parameter separately
        let best_of = self.validation.validate_best_of(best_of)?;

        // create multiple generate requests
        let mut infer_responses: Vec<InferResponse> =
            try_join_all((0..best_of).map(|_| self.generate(request.clone()))).await?;

        // get the sequence with the highest log probability per token
        let mut max_index = 0;
        let mut max_logprob: f32 = f32::MIN;

        for (i, response) in infer_responses.iter().enumerate() {
            // mean logprobs of the generated tokens
            let sequence_logprob = response
                .tokens
                .iter()
                .map(|token| token.logprob)
                .sum::<f32>()
                / response.tokens.len() as f32;

            // set best sequence
            if sequence_logprob > max_logprob {
                max_index = i;
                max_logprob = sequence_logprob;
            }
        }
        let best_response = infer_responses.remove(max_index);
        Ok((best_response, infer_responses))
    }
237
238
239
240
241
242
243
244
}

/// Batching logic
/// Will be launched in a background Tokio task
///
/// Batches requests and sends them to the inference server
async fn batching_task(
    mut client: ShardedClient,
245
246
    waiting_served_ratio: f32,
    max_batch_total_tokens: u32,
247
    max_waiting_tokens: usize,
248
    queue: Queue,
249
    shared: Arc<Shared>,
250
    generation_health: Arc<AtomicBool>,
251
252
253
254
255
256
) {
    // Infinite loop
    loop {
        // Wait for a notification from the Infer struct
        shared.batching_task.notified().await;

257
        // Get the next batch from the queue
258
        // This batch might be smaller than the maximum batch size if there are not enough requests
259
        // waiting in the queue
260
261
262
        while let Some((mut entries, batch, span)) =
            queue.next_batch(None, max_batch_total_tokens).await
        {
263
            let mut cached_batch = prefill(&mut client, batch, &mut entries, &generation_health)
264
265
                .instrument(span)
                .await;
266
267
268
269
270
271
272
            let mut waiting_tokens = 1;

            // We loop until we do not receive any cached batch from the inference server (== until
            // all requests have met their stopping criteria)
            while let Some(batch) = cached_batch {
                // Get current batch info
                let batch_size = batch.size;
273
                let batch_max_tokens = batch.max_tokens;
274
                let mut batches = vec![batch];
275
                metrics::gauge!("tgi_batch_current_size", batch_size as f64);
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
                metrics::gauge!("tgi_batch_current_max_tokens", batch_max_tokens as f64);

                let min_size = if waiting_tokens >= max_waiting_tokens {
                    // If we didn't onboard any new requests since >= max_waiting_tokens, we try
                    // to add a new batch even though its size might be small
                    None
                } else {
                    // Minimum batch size
                    Some((batch_size as f32 * waiting_served_ratio).floor() as usize)
                };

                let token_budget = max_batch_total_tokens - batch_max_tokens;

                // Try to get a new batch
                if let Some((mut new_entries, new_batch, span)) =
                    queue.next_batch(min_size, token_budget).await
                {
                    // Tracking metrics
                    if min_size.is_some() {
                        metrics::increment_counter!("tgi_batch_concat", "reason" => "backpressure");
                    } else {
                        metrics::increment_counter!("tgi_batch_concat", "reason" => "wait_exceeded");
                    }
299

300
301
302
303
304
305
306
307
308
309
310
311
                    entries.iter_mut().for_each(|(_, entry)| {
                        // Create a new span to add the info that this entry is waiting
                        // because a new batch is being computed
                        let entry_waiting_span = info_span!(parent: &entry.span, "waiting");
                        // Add relationships
                        span.follows_from(&entry_waiting_span);
                        entry_waiting_span.follows_from(&span);
                        // Update entry
                        entry.temp_span = Some(entry_waiting_span);
                    });

                    // Generate one token for this new batch to have the attention past in cache
312
313
314
315
                    let new_cached_batch =
                        prefill(&mut client, new_batch, &mut new_entries, &generation_health)
                            .instrument(span)
                            .await;
316
317
318
319
320
321
                    // Reset waiting counter
                    waiting_tokens = 1;
                    // Extend current batch with the new batch
                    if let Some(new_cached_batch) = new_cached_batch {
                        entries.extend(new_entries);
                        batches.push(new_cached_batch);
322
323
                    }
                }
324

325
326
327
328
329
330
                // Create span for this batch to add context to inference calls
                let next_batch_size = entries.len();
                let next_batch_span =
                    info_span!(parent: None, "batch", batch_size = next_batch_size);
                entries.iter_mut().for_each(|(_, entry)| {
                    // Create a new span to link the batch back to this entry
331
                    let entry_batch_span = info_span!(parent: &entry.span, "infer");
332
333
                    // Add relationships
                    next_batch_span.follows_from(&entry_batch_span);
334
335
336
337
                    entry_batch_span.follows_from(&next_batch_span);
                    // Update entry
                    entry.temp_span = Some(entry_batch_span);
                });
338

339
                cached_batch = decode(&mut client, batches, &mut entries, &generation_health)
340
341
                    .instrument(next_batch_span)
                    .await;
342
343
                waiting_tokens += 1;
            }
344
            metrics::gauge!("tgi_batch_current_size", 0.0);
345
            metrics::gauge!("tgi_batch_current_max_tokens", 0.0);
346
347
348
349
        }
    }
}

350
#[instrument(skip_all)]
351
352
353
async fn prefill(
    client: &mut ShardedClient,
    batch: Batch,
354
    entries: &mut IntMap<u64, Entry>,
355
    generation_health: &Arc<AtomicBool>,
356
) -> Option<CachedBatch> {
357
    let start_time = Instant::now();
358
    let batch_id = batch.id;
359
    metrics::increment_counter!("tgi_batch_inference_count", "method" => "prefill");
360
361
362

    match client.prefill(batch).await {
        Ok((generations, next_batch)) => {
363
364
            // Update health
            generation_health.store(true, Ordering::SeqCst);
365
            // Send generated tokens and filter stopped entries
366
367
368
            filter_send_generations(generations, entries);

            // Filter next batch and remove requests that were stopped
369
            let next_batch = filter_batch(client, next_batch, entries).await;
370

371
            metrics::histogram!("tgi_batch_inference_duration", start_time.elapsed().as_secs_f64(), "method" => "prefill");
372
373
374
375
376
            metrics::increment_counter!("tgi_batch_inference_success", "method" => "prefill");
            next_batch
        }
        // If we have an error, we discard the whole batch
        Err(err) => {
377
378
            // Update health
            generation_health.store(false, Ordering::SeqCst);
379
            let _ = client.clear_cache(Some(batch_id)).await;
380
381
382
383
384
385
386
387
388
389
            send_errors(err, entries);
            metrics::increment_counter!("tgi_batch_inference_failure", "method" => "prefill");
            None
        }
    }
}

#[instrument(skip_all)]
async fn decode(
    client: &mut ShardedClient,
390
    batches: Vec<CachedBatch>,
391
    entries: &mut IntMap<u64, Entry>,
392
    generation_health: &Arc<AtomicBool>,
393
) -> Option<CachedBatch> {
394
    let start_time = Instant::now();
395
    let batch_ids: Vec<u64> = batches.iter().map(|b| b.id).collect();
396
    metrics::increment_counter!("tgi_batch_inference_count", "method" => "decode");
397
398

    match client.decode(batches).await {
399
        Ok((generations, next_batch)) => {
400
401
            // Update health
            generation_health.store(true, Ordering::SeqCst);
402
            // Send generated tokens and filter stopped entries
403
404
405
            filter_send_generations(generations, entries);

            // Filter next batch and remove requests that were stopped
406
            let next_batch = filter_batch(client, next_batch, entries).await;
407

408
            metrics::histogram!("tgi_batch_inference_duration", start_time.elapsed().as_secs_f64(), "method" => "decode");
409
            metrics::increment_counter!("tgi_batch_inference_success", "method" => "decode");
410
411
412
413
            next_batch
        }
        // If we have an error, we discard the whole batch
        Err(err) => {
414
            generation_health.store(false, Ordering::SeqCst);
415
416
417
            for id in batch_ids {
                let _ = client.clear_cache(Some(id)).await;
            }
418
            send_errors(err, entries);
419
            metrics::increment_counter!("tgi_batch_inference_failure", "method" => "decode");
420
421
422
423
424
            None
        }
    }
}

425
426
/// Filter a `batch` and remove all requests not present in `entries`
#[instrument(skip_all)]
427
428
async fn filter_batch(
    client: &mut ShardedClient,
429
    next_batch: Option<CachedBatch>,
430
    entries: &IntMap<u64, Entry>,
431
) -> Option<CachedBatch> {
432
433
434
435
436
437
438
439
440
441
    let mut batch = next_batch?;

    // No need to filter
    if batch.size as usize == entries.len() {
        return Some(batch);
    }

    let id = batch.id;

    // Retain only requests that are still in entries
442
    batch.request_ids.retain(|id| entries.contains_key(id));
443

444
    if batch.request_ids.is_empty() {
445
446
447
448
449
450
451
452
453
        // All requests have been filtered out
        // Next batch is now empty
        // Clear it from the Python shards cache
        // We unwrap here as we need to panic since we cannot recover if this method fails
        client.clear_cache(Some(id)).await.unwrap();
        None
    } else {
        // Filter Python shard cache
        // We unwrap here as we need to panic since we cannot recover if this method fails
454
        client.filter_batch(id, batch.request_ids).await.unwrap()
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
    }
}

/// Send one or multiple `InferStreamResponse` to Infer for all `entries`
/// and filter entries
#[instrument(skip_all)]
fn filter_send_generations(generations: Vec<Generation>, entries: &mut IntMap<u64, Entry>) {
    generations.into_iter().for_each(|generation| {
        let id = generation.request_id;
        // Get entry
        // We can `expect` here as the request id should always be in the entries
        let entry = entries
            .get(&id)
            .expect("ID not found in entries. This is a bug.");

        // Create and enter a span to link this function back to the entry
        let _span = info_span!(parent: entry.temp_span.as_ref().expect("batch_span is None. This is a bug."), "send_generation", generation = ?generation).entered();
        // Send generation responses back to the infer task
        // If the receive an error from the Flume channel, it means that the client dropped the
        // request and we need to stop generating hence why we unwrap_or(true)
        let stopped = send_responses(generation, entry).map_err(|err| {
476
477
478
479
            if let SendTimeoutError::Timeout(_) = *err {
                tracing::error!("Entry response channel timed out.")
            }

480
481
482
483
484
485
486
487
488
489
490
491
492
            metrics::increment_counter!("tgi_request_failure", "err" => "dropped");
            err
        }).unwrap_or(true);
        if stopped {
            entries.remove(&id).expect("ID not found in entries. This is a bug.");
        }
    });
}

/// Send responses through the `entry` response channel
fn send_responses(
    generation: Generation,
    entry: &Entry,
493
494
495
496
497
498
) -> Result<bool, Box<SendTimeoutError<Result<InferStreamResponse, InferError>>>> {
    // Return directly if the channel is disconnected
    if entry.response_tx.is_disconnected() {
        return Ok(true);
    }

499
500
501
502
    let mut stopped = false;

    if let Some(prefill_tokens) = generation.prefill_tokens {
        // Send message
503
504
505
506
        entry.response_tx.send_timeout(
            Ok(InferStreamResponse::Prefill(prefill_tokens)),
            Duration::from_millis(10),
        )?;
507
508
509
510
511
512
513
514
515
516
517
518
519
520
    }

    // Create last Token
    let token = Token {
        id: generation.token_id,
        text: generation.token_text,
        logprob: generation.token_logprob,
        special: generation.token_is_special,
    };

    if let Some(generated_text) = generation.generated_text {
        // Generation has ended
        stopped = true;
        // Send message
521
522
523
524
525
526
527
528
529
        entry.response_tx.send_timeout(
            Ok(InferStreamResponse::End {
                token,
                generated_text,
                queued: entry.queue_time,
                start: entry.batch_time.unwrap(),
            }),
            Duration::from_millis(10),
        )?;
530
531
    } else {
        // Send message
532
533
534
535
        entry.response_tx.send_timeout(
            Ok(InferStreamResponse::Token(token)),
            Duration::from_millis(10),
        )?;
536
537
538
539
    }
    Ok(stopped)
}

540
/// Send errors to Infer for all `entries`
541
542
#[instrument(skip_all)]
fn send_errors(error: ClientError, entries: &mut IntMap<u64, Entry>) {
543
    entries.drain().for_each(|(_, entry)| {
544
545
546
        // Create and enter a span to link this function back to the entry
        let _send_error_span = info_span!(parent: entry.temp_span.as_ref().expect("batch_span is None. This is a bug."), "send_error").entered();
        let err = InferError::GenerationError(error.to_string());
547
        metrics::increment_counter!("tgi_request_failure", "err" => "generation");
548
549
        tracing::error!("{err}");

550
551
552
        // unwrap_or is valid here as we don't care if the receiver is gone.
        entry
            .response_tx
553
            .send_timeout(Err(err), Duration::from_millis(10))
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
            .unwrap_or(());
    });
}

#[derive(Debug)]
pub(crate) enum InferStreamResponse {
    // Optional first message
    Prefill(PrefillTokens),
    // Intermediate messages
    Token(Token),
    // Last message
    End {
        token: Token,
        generated_text: GeneratedText,
        start: Instant,
        queued: Instant,
    },
}

#[derive(Debug)]
pub(crate) struct InferResponse {
575
    pub(crate) prefill: Vec<PrefillToken>,
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
    pub(crate) tokens: Vec<Token>,
    pub(crate) generated_text: GeneratedText,
    pub(crate) queued: Instant,
    pub(crate) start: Instant,
}

#[derive(Debug, Error)]
pub enum InferError {
    #[error("Request failed during generation: {0}")]
    GenerationError(String),
    #[error("Model is overloaded")]
    Overloaded(#[from] TryAcquireError),
    #[error("Input validation error: {0}")]
    ValidationError(#[from] ValidationError),
    #[error("Incomplete generation")]
    IncompleteGeneration,
}
593
594
595
596
597
598
599
600
601
602
603

impl InferError {
    pub(crate) fn error_type(&self) -> &str {
        match self {
            InferError::GenerationError(_) => "generation",
            InferError::Overloaded(_) => "overloaded",
            InferError::ValidationError(_) => "validation",
            InferError::IncompleteGeneration => "incomplete_generation",
        }
    }
}