sequence.rs 16.3 KB
Newer Older
1
// SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
// SPDX-License-Identifier: Apache-2.0

4
use crate::protocols::MoveBlock;
5
use derive_getters::Getters;
6
7
use dynamo_tokens::blocks::UniqueBlock;
use dynamo_tokens::{TokenBlockSequence, Tokens};
8
9
10
11
12
use rand::random;

/// Create unique blocks from a TokenBlockSequence
fn create_unique_blocks_from_sequence(
    tokens: &TokenBlockSequence,
13
14
    block_size: usize,
    enable_prefix_caching: bool,
15
16
17
18
) -> Vec<UniqueBlock> {
    let mut unique_blocks: Vec<UniqueBlock> = tokens
        .blocks()
        .iter()
19
20
21
22
23
24
25
        .map(|block| {
            if enable_prefix_caching {
                UniqueBlock::FullBlock(block.sequence_hash())
            } else {
                UniqueBlock::FullBlock(random::<u64>())
            }
        })
26
27
28
        .collect();

    // Only push the partial block if tokens count isn't a multiple of block_size
29
    if !tokens.total_tokens().is_multiple_of(block_size) {
Yan Ru Pei's avatar
Yan Ru Pei committed
30
        unique_blocks.push(UniqueBlock::default());
31
32
33
34
35
36
37
38
39
40
41
42
43
    }
    unique_blocks
}

/// A sequence that is actively being built, with the ability to add tokens and commit to hashes
/// TODO: reuse tokens
#[derive(Debug, Getters)]
pub struct ActiveSequence {
    unique_blocks: Vec<UniqueBlock>,

    tokens: TokenBlockSequence,

    #[getter(copy)]
44
    block_size: usize,
45
46
47
48
49
50
51
52
53
54
55

    #[getter(copy)]
    max_output_tokens: usize,

    #[getter(copy)]
    generated_tokens: usize,

    #[getter(copy)]
    num_input_tokens: usize,

    creation_signal: Option<MoveBlock>,
56
57
58

    #[getter(copy)]
    enable_prefix_caching: bool,
59
60
61
62
63
64
65
}

impl ActiveSequence {
    /// Create a new ActiveSequence instance with the provided tokens
    pub fn new(
        tokens: Vec<u32>,
        max_output_tokens: usize,
66
67
        block_size: Option<usize>,
        enable_prefix_caching: bool,
68
69
70
71
72
    ) -> Self {
        let block_size = block_size.unwrap_or(64);
        assert!(block_size > 1, "block_size must be greater than 1");
        let num_input_tokens = tokens.len();

73
        let tokens = Tokens::from(tokens).into_sequence(block_size as u32, Some(1337));
74
        let unique_blocks =
Yan Ru Pei's avatar
Yan Ru Pei committed
75
76
77
            create_unique_blocks_from_sequence(&tokens, block_size, enable_prefix_caching);
        let block_hashes = tokens.blocks().iter().map(|b| b.block_hash()).collect();
        let creation_signal = Some(MoveBlock::Use(unique_blocks.clone(), block_hashes));
78
79
80
81
82
83
84
85
86

        Self {
            unique_blocks,
            tokens,
            block_size,
            max_output_tokens,
            generated_tokens: 0,
            num_input_tokens,
            creation_signal,
87
            enable_prefix_caching,
88
89
90
        }
    }

91
    pub fn extra_tokens(&self) -> u32 {
92
        (self.len() % self.block_size) as u32
93
94
95
96
97
98
99
100
101
102
    }

    pub fn len(&self) -> usize {
        self.tokens.total_tokens()
    }

    pub fn is_empty(&self) -> bool {
        self.tokens.total_tokens() == 0
    }

103
104
105
106
    pub fn take_creation_signal(&mut self) -> Option<MoveBlock> {
        self.creation_signal.take()
    }

107
108
109
110
111
112
113
114
    pub fn block_hashes(&self) -> Vec<u64> {
        self.tokens
            .blocks()
            .iter()
            .map(|block| block.block_hash())
            .collect()
    }

115
116
117
118
    /// Create a new ActiveSequence instance and return the creation signal
    pub fn new_with_signal(
        tokens: Vec<u32>,
        max_output_tokens: usize,
119
120
        block_size: Option<usize>,
        enable_prefix_caching: bool,
121
    ) -> (Self, Option<MoveBlock>) {
122
        let mut sequence = Self::new(tokens, max_output_tokens, block_size, enable_prefix_caching);
123
        let signal = sequence.take_creation_signal();
124
125
126
127
128
129
130
131
        (sequence, signal)
    }

    /// Push a token to the sequence
    pub fn push(&mut self, token: u32) -> Option<Vec<MoveBlock>> {
        self.tokens.append(token).expect("Token push failed.");
        self.generated_tokens += 1;

132
        if self.len() % self.block_size != 1 {
133
134
135
136
137
138
139
140
141
            return None;
        }

        // Add a partial block for the first token in a new partial sequence
        // Send Use signal (to allocate space for this new generation block)
        let mut signals = Vec::new();

        // Replace last partial block with full block if it exists
        if let Some(UniqueBlock::PartialBlock(uuid)) = self.unique_blocks.last().cloned() {
Yan Ru Pei's avatar
Yan Ru Pei committed
142
            let last_seq_hash = if self.enable_prefix_caching {
143
144
145
146
                self.tokens.last_complete_block().unwrap().sequence_hash()
            } else {
                random::<u64>()
            };
Yan Ru Pei's avatar
Yan Ru Pei committed
147
            let last_block_hash = self.tokens.last_complete_block().unwrap().block_hash();
148
            self.unique_blocks.pop();
Yan Ru Pei's avatar
Yan Ru Pei committed
149
150
151
152
153
154
155

            // After pop, the last element is the parent block
            let second_to_last_hash = self.unique_blocks.last().map(|block| match block {
                UniqueBlock::FullBlock(hash) => *hash,
                UniqueBlock::PartialBlock(_) => panic!("Cannot have a partial block as parent"),
            });

156
            self.unique_blocks
Yan Ru Pei's avatar
Yan Ru Pei committed
157
                .push(UniqueBlock::FullBlock(last_seq_hash));
158
159
            signals.push(MoveBlock::Promote(
                uuid,
Yan Ru Pei's avatar
Yan Ru Pei committed
160
161
                last_seq_hash,
                second_to_last_hash,
162
163
                last_block_hash,
            ));
164
165
166
167
        }

        let new_partial_block = UniqueBlock::default();
        self.unique_blocks.push(new_partial_block.clone());
Yan Ru Pei's avatar
Yan Ru Pei committed
168
        signals.push(MoveBlock::Use(vec![new_partial_block], vec![]));
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
        Some(signals)
    }

    /// Generate a random token, push it to the sequence, and increment generation count.
    ///
    /// This function:
    /// - Generates a random token and adds it to the current sequence
    /// - Acquires a new partial block if needed or promotes an existing partial block to a full block
    /// - Returns appropriate signals for the KvManager to process
    ///
    /// # Panics
    ///
    /// Calling this function when max_output_tokens has already been reached will cause a panic.
    /// Always check `generated_tokens < max_output_tokens` before calling this method.
    pub fn generate(&mut self) -> Vec<MoveBlock> {
        // Assert that we haven't reached the maximum output tokens
        assert!(
            self.generated_tokens < self.max_output_tokens,
            "Cannot generate more tokens: reached max_output_tokens limit"
        );

        // Generate a random token
        let token = random::<u32>();

        // Collect signals
        let mut signals = Vec::new();

        // Push the token to the sequence and collect any signals
        if let Some(move_blocks) = self.push(token) {
            signals.extend(move_blocks);
        }

        // Check if we've reached the limit after pushing
        if self.generated_tokens != self.max_output_tokens {
            return signals;
        }

        // Free all blocks when we reach max tokens
        signals.extend(self.free_signal());
        signals
    }

    /// Free all blocks, generating appropriate signals for each block type
    pub fn free_signal(&self) -> Vec<MoveBlock> {
        self.unique_blocks
            .iter()
            .rev()
            .map(|block| match block {
                UniqueBlock::PartialBlock(uuid) => {
                    MoveBlock::Destroy(vec![UniqueBlock::PartialBlock(*uuid)])
                }
                UniqueBlock::FullBlock(hash) => {
                    MoveBlock::Deref(vec![UniqueBlock::FullBlock(*hash)])
                }
            })
            .collect()
    }

227
228
    /// Move the request to a preempted state and return the free signals from freeing current blocks
    /// Upon preemption, the sequence retains the tokens generated during the decode phase (if any).
229
230
231
    pub fn reset_with_signal(&mut self) -> Vec<MoveBlock> {
        let free_signal = self.free_signal();

232
233
        // Don't reset generated_tokens since we're keeping the tokens in the sequence

Yan Ru Pei's avatar
Yan Ru Pei committed
234
235
236
237
        self.creation_signal = Some(MoveBlock::Use(
            self.unique_blocks.clone(),
            self.block_hashes(),
        ));
238
239
240
241
242
243
244
245
246
247

        free_signal
    }

    /// Pops last token in the sequence.
    pub fn pop(&mut self) {
        self.tokens.pop();
        self.generated_tokens = self.generated_tokens.saturating_sub(1);

        // Reverts to the last full block
248
        if self.tokens.total_tokens().is_multiple_of(self.block_size) {
249
250
251
252
253
254
255
256
257
258
259
260
261
262
            self.unique_blocks.pop();
        }
    }
}

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

    #[test]
    fn test_active_sequence_push() {
        // Create a sequence with block size 16 initialized with tokens [0..15]
        let initial_tokens: Vec<u32> = (0..15).collect();
        let (mut seq1, signal1) =
263
            ActiveSequence::new_with_signal(initial_tokens, 100, Some(16), true);
264
265
266
267
268
269
        assert_eq!(seq1.num_input_tokens(), 15);
        assert_eq!(seq1.len(), 15);

        // Check that we got a Use signal
        assert!(signal1.is_some());
        match &signal1 {
Yan Ru Pei's avatar
Yan Ru Pei committed
270
            Some(MoveBlock::Use(blocks, _hashes)) => {
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
                assert_eq!(blocks.len(), 1);
            }
            _ => panic!("Expected Use signal"),
        }

        // Push token 15 which should complete the block (no signals yet)
        let signal_15 = seq1.push(15);
        assert!(
            signal_15.is_none(),
            "Completing a block should not trigger signals"
        );

        // Push token 16 which should trigger both Promote and Use signals
        let signal_16 = seq1.push(16);
        assert!(signal_16.is_some());
        let signal_16 = signal_16.unwrap();
        assert_eq!(signal_16.len(), 2);

289
290
        // First signal should be Promote for the previous block
        match &signal_16[0] {
Yan Ru Pei's avatar
Yan Ru Pei committed
291
            MoveBlock::Promote(_, _, parent_hash, _hash) => {
292
293
294
295
296
                assert_eq!(*parent_hash, None);
            }
            _ => panic!("Expected Promote signal as second signal"),
        }

297
298
        // Second signal should be Use for new partial block
        match &signal_16[1] {
Yan Ru Pei's avatar
Yan Ru Pei committed
299
            MoveBlock::Use(blocks, _hashes) => {
300
301
302
303
304
305
306
307
308
                assert_eq!(blocks.len(), 1);
                assert!(matches!(blocks[0], UniqueBlock::PartialBlock(_)));
            }
            _ => panic!("Expected Use signal as first signal"),
        }

        // Verify state after pushing tokens
        assert_eq!(seq1.unique_blocks().len(), 2); // One full block and one partial block
        assert_eq!(seq1.len(), 17);
309
        assert_eq!(seq1.len() % seq1.block_size(), 1);
310
311
312

        // Create another sequence with block size 16 initialized with tokens [0..17]
        let extended_tokens: Vec<u32> = (0..16).collect();
313
        let (mut seq2, _) = ActiveSequence::new_with_signal(extended_tokens, 100, Some(16), true);
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
        seq2.push(16);
        seq2.pop();
        seq2.push(16);

        // Simplified assertions
        assert_eq!(
            seq1.unique_blocks()[0],
            seq2.unique_blocks()[0],
            "First blocks should be the same"
        );

        assert_ne!(
            seq1.unique_blocks()[1],
            seq2.unique_blocks()[1],
            "Second blocks should be different"
        );

        // Reset partial block on seq1 and push back token 16
        seq1.push(17);
        seq1.pop();
        seq1.pop();
        seq1.push(16);

        // Now push tokens 17..32 to both sequences
        for token in 17..33 {
            seq1.push(token);
            seq2.push(token);
        }

        // Both sequences should now have 2 blocks:
        // 1. FullBlock for tokens 0-15
        // 2. FullBlock for tokens 16-31
        // 3. No partial block since there are no remaining tokens
        assert_eq!(
            seq1.unique_blocks().len(),
            3,
            "seq1 should have exactly 3 blocks"
        );
        assert_eq!(
            seq2.unique_blocks().len(),
            3,
            "seq2 should have exactly 3 blocks"
        );
        assert_eq!(
358
            seq1.len() % seq1.block_size(),
359
360
361
362
            1,
            "seq1 should have 1 partial token"
        );
        assert_eq!(
363
            seq2.len() % seq2.block_size(),
364
365
366
367
368
369
370
371
372
373
374
            1,
            "seq2 should have 1 partial token"
        );

        // Verify that both sequences have identical blocks up to the second position
        assert_eq!(
            &seq1.unique_blocks()[0..2],
            &seq2.unique_blocks()[0..2],
            "First two blocks should be identical"
        );

375
376
377
378
379
380
381
382
383
384
385
        // Push tokens 34..47 to seq1
        for token in 33..48 {
            seq1.push(token);
        }

        // Push token 48 and get the signal - this completes the block and triggers signals
        let signal = seq1.push(48);
        let signal = signal.unwrap();

        // Check that signal[0] is promote
        match &signal[0] {
Yan Ru Pei's avatar
Yan Ru Pei committed
386
            MoveBlock::Promote(_, _, parent_hash, _hash) => {
387
388
389
390
391
392
393
394
395
396
397
398
399
400
                // Check that the parent_hash matches unique_blocks[1], which should be a full block
                if let UniqueBlock::FullBlock(expected_hash) = seq1.unique_blocks()[1] {
                    assert_eq!(
                        *parent_hash,
                        Some(expected_hash),
                        "Parent hash should match unique_blocks[1]"
                    );
                } else {
                    panic!("unique_blocks[1] should be a full block");
                }
            }
            _ => panic!("Expected Promote signal as first signal"),
        }

401
402
403
        // Reset seq1 and check that it equals the original clone
        let free_signals = seq1.reset_with_signal();

404
        // 49 - 15 generated tokens
405
        assert_eq!(seq1.generated_tokens(), 34);
406

407
408
409
410
411
412
413
414
        // Verify the reset signals include proper cleanup events
        assert!(!free_signals.is_empty());
    }

    #[test]
    fn test_active_sequence_generate_signals() {
        // Create a sequence with block size 16, max_output_tokens 4, initialized with tokens [0..14)
        let initial_tokens: Vec<u32> = (0..14).collect();
415
        let (mut seq, signal) = ActiveSequence::new_with_signal(initial_tokens, 5, Some(16), true);
416
417
418
419

        // Initial signal - should have received a Use signal for the partial block
        assert!(signal.is_some());
        match signal {
Yan Ru Pei's avatar
Yan Ru Pei committed
420
            Some(MoveBlock::Use(blocks, _hashes)) => {
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
                assert_eq!(blocks.len(), 1);
                assert!(matches!(blocks[0], UniqueBlock::PartialBlock(_)));
            }
            _ => panic!("Expected Use signal for the initial partial block"),
        }

        // Generate first two tokens - should not trigger new signals
        seq.generate();
        let signals_first = seq.generate();
        assert_eq!(signals_first.len(), 0);

        // Generate third token - this fills the block and should trigger both Promote and Use signals
        let signals_second = seq.generate();
        assert_eq!(signals_second.len(), 2);

436
437
        // First signal should be Promote
        match &signals_second[0] {
Yan Ru Pei's avatar
Yan Ru Pei committed
438
            MoveBlock::Promote(_, _, parent_hash, _hash) => {
439
440
441
442
443
444
                assert_eq!(*parent_hash, None);
            }
            _ => panic!("Expected Promote signal as first signal after second token"),
        }

        // Second signal should be Use for new partial block
445
        match &signals_second[1] {
Yan Ru Pei's avatar
Yan Ru Pei committed
446
            MoveBlock::Use(blocks, _hashes) => {
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
                assert_eq!(blocks.len(), 1);
                assert!(matches!(blocks[0], UniqueBlock::PartialBlock(_)));
            }
            _ => panic!("Expected Use signal as second signal after second token"),
        }

        // Generate fourth token - should not trigger new signals as it's adding to partial block
        let signals_third = seq.generate();
        assert_eq!(signals_third.len(), 0);

        // Generate last token - we reach max_output_tokens, should trigger Destroy and Deref signals
        let signals_last = seq.generate();
        assert_eq!(signals_last.len(), 2);

        // First signal should be Destroy for the partial block
        match &signals_last[0] {
            MoveBlock::Destroy(blocks) => {
                assert_eq!(blocks.len(), 1);
                assert!(matches!(blocks[0], UniqueBlock::PartialBlock(_)));
            }
            _ => panic!("Expected Destroy signal for partial block after fourth token"),
        }

        // Second signal should be Deref for the full block
        match &signals_last[1] {
            MoveBlock::Deref(blocks) => {
                assert_eq!(blocks.len(), 1);
                assert!(matches!(blocks[0], UniqueBlock::FullBlock(_)));
            }
            _ => panic!("Expected Deref signal for full block after fourth token"),
        }
    }
}