spec-stream-engine 0.1.0

Speculative decoding async execution engine, pipeline scheduler, and token streaming orchestrator
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
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
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
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
use crate::error::EngineError;
use crate::model::ModelEngine;
use crate::pipeline::SpeculativeConfig;
use crate::stream::GenerationEvent;
use spec_stream_cache::PagedKvCacheManager;
use spec_stream_core::{DraftProposal, Logits, SpeculativeMetrics, SpeculativeVerifier, TokenId};
use std::collections::VecDeque;
use std::sync::atomic::{AtomicUsize, Ordering};
use std::sync::{Arc, Mutex as StdMutex};
use tokio::sync::{mpsc, Mutex as AsyncMutex};
use tokio::time::Instant;

/// An incoming inference request submitted to the batch scheduler.
pub struct QueuedRequest {
    pub sequence_id: u64,
    pub prompt: Vec<TokenId>,
    pub config: SpeculativeConfig,
    pub sender: mpsc::Sender<GenerationEvent>,
    pub priority: usize,
    pub created_at: Instant,
}

impl QueuedRequest {
    pub fn new(
        sequence_id: u64,
        prompt: Vec<TokenId>,
        config: SpeculativeConfig,
        sender: mpsc::Sender<GenerationEvent>,
    ) -> Self {
        Self {
            sequence_id,
            prompt,
            config,
            sender,
            priority: 0,
            created_at: Instant::now(),
        }
    }
}

/// An active in-flight sequence currently participating in iteration-level batching.
struct ActiveSequence {
    sequence_id: u64,
    current_prefix: Vec<TokenId>,
    total_generated: usize,
    config: SpeculativeConfig,
    current_k: usize,
    ema_acceptance: Option<f32>,
    metrics: SpeculativeMetrics,
    sender: mpsc::Sender<GenerationEvent>,
}

/// Real-time operational statistics of the continuous batch scheduler.
#[derive(Debug, Clone, Default)]
pub struct SchedulerStats {
    pub active_sequences: usize,
    pub queued_requests: usize,
    pub completed_sequences: usize,
    pub total_tokens_emitted: usize,
    pub total_batch_steps: usize,
}

/// High-throughput continuous batch scheduler executing iteration-level speculative inference.
/// Dynamically pools multiple concurrent requests into unified forward passes.
pub struct ContinuousBatchScheduler<D: ModelEngine, T: ModelEngine> {
    draft_model: Arc<AsyncMutex<D>>,
    target_model: Arc<AsyncMutex<T>>,
    cache_manager: Arc<PagedKvCacheManager>,
    queue: Arc<StdMutex<VecDeque<QueuedRequest>>>,
    active: Arc<StdMutex<Vec<ActiveSequence>>>,
    max_batch_size: usize,
    completed_counter: Arc<AtomicUsize>,
    tokens_emitted_counter: Arc<AtomicUsize>,
    batch_steps_counter: Arc<AtomicUsize>,
}

impl<D: ModelEngine + 'static, T: ModelEngine + 'static> ContinuousBatchScheduler<D, T> {
    pub fn new(
        draft_model: D,
        target_model: T,
        cache_manager: Arc<PagedKvCacheManager>,
        max_batch_size: usize,
    ) -> Self {
        Self {
            draft_model: Arc::new(AsyncMutex::new(draft_model)),
            target_model: Arc::new(AsyncMutex::new(target_model)),
            cache_manager,
            queue: Arc::new(StdMutex::new(VecDeque::new())),
            active: Arc::new(StdMutex::new(Vec::new())),
            max_batch_size: max_batch_size.max(1),
            completed_counter: Arc::new(AtomicUsize::new(0)),
            tokens_emitted_counter: Arc::new(AtomicUsize::new(0)),
            batch_steps_counter: Arc::new(AtomicUsize::new(0)),
        }
    }

    /// Enqueues an inference request for continuous batch processing.
    pub fn submit(&self, request: QueuedRequest) -> Result<(), EngineError> {
        let mut queue = self.queue.lock().unwrap();
        queue.push_back(request);
        Ok(())
    }

    /// Number of requests currently waiting in the arrival queue.
    pub fn queue_len(&self) -> usize {
        self.queue.lock().unwrap().len()
    }

    /// Number of requests currently active in the generation batch.
    pub fn active_len(&self) -> usize {
        self.active.lock().unwrap().len()
    }

    /// Retrieves current scheduler telemetry.
    pub fn stats(&self) -> SchedulerStats {
        SchedulerStats {
            active_sequences: self.active.lock().unwrap().len(),
            queued_requests: self.queue.lock().unwrap().len(),
            completed_sequences: self.completed_counter.load(Ordering::Relaxed),
            total_tokens_emitted: self.tokens_emitted_counter.load(Ordering::Relaxed),
            total_batch_steps: self.batch_steps_counter.load(Ordering::Relaxed),
        }
    }

    /// Executes one iteration of continuous batching across all currently scheduled sequences.
    /// Returns the total number of committed tokens produced across all sequences in this iteration.
    pub async fn step_iteration(&self) -> Result<usize, EngineError> {
        self.batch_steps_counter.fetch_add(1, Ordering::Relaxed);

        // 1. Admission Control: Promote queued requests up to max_batch_size
        {
            let mut active = self.active.lock().unwrap();
            let mut queue = self.queue.lock().unwrap();

            while active.len() < self.max_batch_size && !queue.is_empty() {
                // Check if the cache has sufficient blocks
                if self.cache_manager.free_blocks() < 2 {
                    break;
                }

                if let Some(req) = queue.pop_front() {
                    self.cache_manager.create_sequence(req.sequence_id);
                    let _ = self
                        .cache_manager
                        .append_tokens(req.sequence_id, &req.prompt);

                    let initial_k = req.config.k_draft_tokens;
                    active.push(ActiveSequence {
                        sequence_id: req.sequence_id,
                        current_prefix: req.prompt,
                        total_generated: 0,
                        config: req.config,
                        current_k: initial_k,
                        ema_acceptance: None,
                        metrics: SpeculativeMetrics::new(),
                        sender: req.sender,
                    });
                }
            }
        }

        let num_active = self.active.lock().unwrap().len();
        if num_active == 0 {
            return Ok(0);
        }

        // 2. Draft Proposals Phase (per sequence)
        struct SeqProposal {
            seq_idx: usize,
            last_token: TokenId,
            proposal: DraftProposal,
        }

        let mut proposals = Vec::with_capacity(num_active);
        {
            let active = self.active.lock().unwrap();
            for (idx, seq) in active.iter().enumerate() {
                if let Some(&last_t) = seq.current_prefix.last() {
                    proposals.push(SeqProposal {
                        seq_idx: idx,
                        last_token: last_t,
                        proposal: DraftProposal::new(),
                    });
                }
            }
        }

        // Generate draft tokens
        for p in &mut proposals {
            let k = self.active.lock().unwrap()[p.seq_idx].current_k;
            let mut draft_token = p.last_token;

            for _ in 0..k {
                let logits = {
                    let mut draft_guard = self.draft_model.lock().await;
                    draft_guard.forward_token(draft_token).await?
                };

                let next_id = match logits.argmax() {
                    Ok(id) => id,
                    Err(_) => break,
                };

                let prob = logits
                    .softmax()
                    .ok()
                    .and_then(|probs| probs.get(next_id as usize).copied())
                    .unwrap_or(0.85);

                p.proposal.push(next_id, prob);
                draft_token = next_id;
            }
        }

        // 3. Target Model Batch Forward Pass & Verification
        let mut total_tokens_this_step = 0;
        let mut finished_indices = Vec::new();

        for p in proposals {
            if p.proposal.is_empty() {
                continue;
            }

            let proposed_ids = p.proposal.token_ids();
            let seq_id = self.active.lock().unwrap()[p.seq_idx].sequence_id;
            let _ = self.cache_manager.append_tokens(seq_id, &proposed_ids);

            let mut target_inputs = Vec::with_capacity(proposed_ids.len() + 1);
            target_inputs.push(p.last_token);
            target_inputs.extend_from_slice(&proposed_ids);

            let target_logits: Vec<Logits> = {
                let mut target_guard = self.target_model.lock().await;
                target_guard.forward_batch(&target_inputs).await?
            };

            let verifier = {
                let active = self.active.lock().unwrap();
                SpeculativeVerifier::new(active[p.seq_idx].config.strategy)
            };

            let verif_result = match verifier.verify(&p.proposal, &target_logits) {
                Ok(r) => r,
                Err(e) => {
                    tracing::error!("Scheduler verification error: {:?}", e);
                    continue;
                }
            };

            // 4. KV-Cache Rollback & Commit in O(1)
            let mut active = self.active.lock().unwrap();
            let seq = &mut active[p.seq_idx];

            let valid_tokens_to_keep = seq.current_prefix.len() + verif_result.accepted_count;
            let _ = self
                .cache_manager
                .rollback_sequence(seq.sequence_id, valid_tokens_to_keep);
            let _ = self
                .cache_manager
                .append_tokens(seq.sequence_id, &[verif_result.emitted_target_token]);

            seq.metrics.record_cycle(
                p.proposal.len(),
                verif_result.accepted_count,
                verif_result.total_tokens_produced,
            );

            // Emit accepted draft tokens
            for &tok in &verif_result.accepted_tokens {
                seq.current_prefix.push(tok);
                seq.total_generated += 1;
                total_tokens_this_step += 1;
                let _ = seq.sender.try_send(GenerationEvent::Token {
                    token_id: tok,
                    is_draft: true,
                });
            }

            // Emit target token
            seq.current_prefix.push(verif_result.emitted_target_token);
            seq.total_generated += 1;
            total_tokens_this_step += 1;
            let _ = seq.sender.try_send(GenerationEvent::Token {
                token_id: verif_result.emitted_target_token,
                is_draft: false,
            });

            // Adaptive K update
            let acc_rate = verif_result.accepted_count as f32 / p.proposal.len().max(1) as f32;
            let ema = match seq.ema_acceptance {
                Some(prev) => 0.5 * acc_rate + 0.5 * prev,
                None => acc_rate,
            };
            seq.ema_acceptance = Some(ema);

            if seq.config.adaptive_k {
                if ema >= 0.75 && seq.current_k < seq.config.max_k {
                    seq.current_k += 1;
                } else if ema <= 0.35 && seq.current_k > seq.config.min_k {
                    seq.current_k -= 1;
                }
            }

            // Check termination conditions
            let eos_reached = seq.config.eos_token_id == Some(verif_result.emitted_target_token);
            let max_reached = seq.total_generated >= seq.config.max_new_tokens;

            if eos_reached || max_reached {
                let reason = if eos_reached { "EOS" } else { "MaxTokens" };
                let _ = seq.sender.try_send(GenerationEvent::Finished {
                    total_tokens: seq.total_generated,
                    reason: reason.into(),
                });
                finished_indices.push(p.seq_idx);
            }
        }

        self.tokens_emitted_counter
            .fetch_add(total_tokens_this_step, Ordering::Relaxed);

        // 5. Retire finished sequences and cleanup their cache
        if !finished_indices.is_empty() {
            finished_indices.sort_unstable_by(|a, b| b.cmp(a)); // remove in reverse order
            let mut active = self.active.lock().unwrap();
            for idx in finished_indices {
                let finished_seq = active.swap_remove(idx);
                let _ = self.cache_manager.remove_sequence(finished_seq.sequence_id);
                self.completed_counter.fetch_add(1, Ordering::Relaxed);
            }
        }

        Ok(total_tokens_this_step)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::model::SyntheticModel;

    #[tokio::test]
    async fn test_continuous_batch_scheduler() {
        let vocab_size = 50;
        let draft = SyntheticModel::new("draft", vocab_size, 0.8, 42);
        let target = SyntheticModel::new("target", vocab_size, 0.95, 1337);
        let cache = Arc::new(PagedKvCacheManager::with_defaults(64));

        let scheduler = ContinuousBatchScheduler::new(draft, target, cache, 4);

        let (tx1, mut rx1) = mpsc::channel(32);
        let (tx2, mut rx2) = mpsc::channel(32);

        let config1 = SpeculativeConfig {
            max_new_tokens: 6,
            ..Default::default()
        };
        let config2 = SpeculativeConfig {
            max_new_tokens: 8,
            ..Default::default()
        };

        scheduler
            .submit(QueuedRequest::new(101, vec![1, 2], config1, tx1))
            .unwrap();
        scheduler
            .submit(QueuedRequest::new(102, vec![3, 4], config2, tx2))
            .unwrap();

        assert_eq!(scheduler.queue_len(), 2);
        assert_eq!(scheduler.active_len(), 0);

        // Run iterations until both complete
        let mut iterations = 0;
        while scheduler.active_len() > 0 || scheduler.queue_len() > 0 {
            let tokens = scheduler.step_iteration().await.unwrap();
            iterations += 1;
            if iterations > 20 {
                break;
            }
            if tokens == 0 && scheduler.active_len() == 0 {
                break;
            }
        }

        let stats = scheduler.stats();
        assert_eq!(stats.completed_sequences, 2);
        assert!(stats.total_tokens_emitted >= 14);

        // Drain channels
        let mut tokens_req1 = 0;
        while let Ok(event) = rx1.try_recv() {
            if let GenerationEvent::Token { .. } = event {
                tokens_req1 += 1;
            }
        }
        assert!(tokens_req1 >= 6);

        let mut tokens_req2 = 0;
        while let Ok(event) = rx2.try_recv() {
            if let GenerationEvent::Token { .. } = event {
                tokens_req2 += 1;
            }
        }
        assert!(tokens_req2 >= 8);
    }
}