scheduler.py 50.4 KB
Newer Older
1
import enum
2
3
import os
import random
4
import time
5
from collections import deque
6
from dataclasses import dataclass, field
7
from typing import Deque, Dict, Iterable, List, Optional, Set, Tuple, Union
Woosuk Kwon's avatar
Woosuk Kwon committed
8

9
from vllm.config import CacheConfig, LoRAConfig, SchedulerConfig
10
from vllm.core.interfaces import AllocStatus, BlockSpaceManager
11
from vllm.core.policy import Policy, PolicyFactory
Woosuk Kwon's avatar
Woosuk Kwon committed
12
from vllm.logger import init_logger
13
from vllm.lora.request import LoRARequest
Woosuk Kwon's avatar
Woosuk Kwon committed
14
from vllm.sequence import (Sequence, SequenceData, SequenceGroup,
15
                           SequenceGroupMetadata, SequenceStatus)
Woosuk Kwon's avatar
Woosuk Kwon committed
16

Woosuk Kwon's avatar
Woosuk Kwon committed
17
logger = init_logger(__name__)
18

19
20
21
22
23
24
25
# Test-only. If configured, decode is preempted with
# ARTIFICIAL_PREEMPTION_PROB% probability.
ENABLE_ARTIFICIAL_PREEMPT = bool(
    os.getenv("VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT", False))  # noqa
ARTIFICIAL_PREEMPTION_PROB = 0.5
ARTIFICIAL_PREEMPTION_MAX_CNT = 500

Woosuk Kwon's avatar
Woosuk Kwon committed
26

27
28
29
30
31
32
33
34
35
36
37
38
39
class PreemptionMode(enum.Enum):
    """Preemption modes.

    1. Swapping: Swap out the blocks of the preempted sequences to CPU memory
    and swap them back in when the sequences are resumed.
    2. Recomputation: Discard the blocks of the preempted sequences and
    recompute them when the sequences are resumed, treating the sequences as
    new prompts.
    """
    SWAP = enum.auto()
    RECOMPUTE = enum.auto()


40
41
@dataclass
class SchedulingBudget:
42
43
44
45
46
47
48
49
50
    """The available slots for scheduling.

    TODO(sang): Right now, the budget is request_id-aware meaning it can ignore
    budget update from the same request_id. It is because in normal scheduling
    path, we update RUNNING num_seqs ahead of time, meaning it could be
    updated more than once when scheduling RUNNING requests. Since this won't
    happen if we only have chunked prefill scheduling, we can remove this
    feature from the API when chunked prefill is enabled by default.
    """
51
52
    token_budget: int
    max_num_seqs: int
53
54
    _requeset_ids_num_batched_tokens: Set[str] = field(default_factory=set)
    _requeset_ids_num_curr_seqs: Set[str] = field(default_factory=set)
55
56
    _num_batched_tokens: int = 0
    _num_curr_seqs: int = 0
57
58

    def can_schedule(self, *, num_new_tokens: int, num_new_seqs: int):
59
60
        assert num_new_tokens != 0
        assert num_new_seqs != 0
61
62
63
        return (self.num_batched_tokens + num_new_tokens <= self.token_budget
                and self.num_curr_seqs + num_new_seqs <= self.max_num_seqs)

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
    def remaining_token_budget(self):
        return self.token_budget - self.num_batched_tokens

    def add_num_batched_tokens(self, req_id: str, num_batched_tokens: int):
        if req_id in self._requeset_ids_num_batched_tokens:
            return

        self._requeset_ids_num_batched_tokens.add(req_id)
        self._num_batched_tokens += num_batched_tokens

    def subtract_num_batched_tokens(self, req_id: str,
                                    num_batched_tokens: int):
        if req_id in self._requeset_ids_num_batched_tokens:
            self._requeset_ids_num_batched_tokens.remove(req_id)
            self._num_batched_tokens -= num_batched_tokens

    def add_num_seqs(self, req_id: str, num_curr_seqs: int):
        if req_id in self._requeset_ids_num_curr_seqs:
            return

        self._requeset_ids_num_curr_seqs.add(req_id)
        self._num_curr_seqs += num_curr_seqs

    def subtract_num_seqs(self, req_id: str, num_curr_seqs: int):
        if req_id in self._requeset_ids_num_curr_seqs:
            self._requeset_ids_num_curr_seqs.remove(req_id)
            self._num_curr_seqs -= num_curr_seqs

    @property
    def num_batched_tokens(self):
        return self._num_batched_tokens

    @property
    def num_curr_seqs(self):
        return self._num_curr_seqs

100

101
102
103
104
105
106
107
108
109
110
@dataclass
class ScheduledSequenceGroup:
    # A sequence group that's scheduled.
    seq_group: SequenceGroup
    # The total chunk size (number of tokens) to process for next iteration.
    # 1 for decoding. Same as prompt tokens for prefill, but if prefill is
    # chunked, it can be smaller than that.
    token_chunk_size: int


111
@dataclass
112
class SchedulerOutputs:
113
    """The scheduling decision made from a scheduler."""
114
115
116
117
118
119
    # Scheduled sequence groups.
    scheduled_seq_groups: Iterable[ScheduledSequenceGroup]
    # Number of prefill groups scheduled.
    num_prefill_groups: int
    # Total number of batched tokens.
    num_batched_tokens: int
120
121
122
123
    # Blocks to swap in. List of CPU -> GPU block number.
    blocks_to_swap_in: List[Tuple[int, int]]
    # Blocks to swap out. List of GPU -> CPU block number.
    blocks_to_swap_out: List[Tuple[int, int]]
124
125
    # Blocks to copy. Source to dest block.
    blocks_to_copy: List[Tuple[int, int]]
126
127
128
129
    # Sequence groups that are going to be ignored.
    ignored_seq_groups: List[SequenceGroup]
    # The number of slots for lookahead decoding.
    num_lookahead_slots: int
130
131
    # The number of requests in the running queue
    running_queue_size: int
132
    preempted: int
133
134

    def __post_init__(self):
135
        # Swap in and swap out should never happen at the same time.
136
        assert not (self.blocks_to_swap_in and self.blocks_to_swap_out)
137

138
        self.num_loras: int = len(self.lora_requests)
139
140
141
        if self.num_loras > 0:
            self._sort_by_lora_ids()

142
    def is_empty(self) -> bool:
Woosuk Kwon's avatar
Woosuk Kwon committed
143
144
145
        # NOTE: We do not consider the ignored sequence groups.
        return (not self.scheduled_seq_groups and not self.blocks_to_swap_in
                and not self.blocks_to_swap_out and not self.blocks_to_copy)
146

147
    def _sort_by_lora_ids(self):
148
149
150
        self.scheduled_seq_groups = sorted(
            self.scheduled_seq_groups,
            key=lambda g: (g.seq_group.lora_int_id, g.seq_group.request_id))
151
152
153

    @property
    def lora_requests(self) -> Set[LoRARequest]:
154
155
156
157
158
        return {
            g.seq_group.lora_request
            for g in self.scheduled_seq_groups
            if g.seq_group.lora_request is not None
        }
159

160

161
@dataclass
162
163
164
165
166
167
168
169
170
171
172
class SchedulerRunningOutputs:
    """The requests that are scheduled from a running queue.

    Could contain prefill (prefill that's chunked) or decodes. If there's not
    enough memory, it can be preempted (for recompute) or swapped out.
    """
    # Selected sequences that are running and in a decoding phase.
    decode_seq_groups: List[SequenceGroup]
    # Selected sequences that are running and in a prefill phase.
    # I.e., it means the prefill has been chunked.
    prefill_seq_groups: List[SequenceGroup]
173
174
175
176
177
    # The preempted sequences.
    preempted: List[SequenceGroup]
    # Sequences that are swapped out.
    swapped_out: List[SequenceGroup]
    # The blocks to swap out.
178
    blocks_to_swap_out: List[Tuple[int, int]]
179
    # The blocks to copy.
180
    blocks_to_copy: List[Tuple[int, int]]
181
    # The number of slots for lookahead decoding.
182
183
184
    num_lookahead_slots: int

    @classmethod
185
186
187
188
    def create_empty(cls) -> "SchedulerRunningOutputs":
        return SchedulerRunningOutputs(
            decode_seq_groups=[],
            prefill_seq_groups=[],
189
190
            preempted=[],
            swapped_out=[],
191
            blocks_to_swap_out=[],
192
            blocks_to_copy=[],
193
194
195
196
197
198
            num_lookahead_slots=0,
        )


@dataclass
class SchedulerSwappedInOutputs:
199
200
201
202
203
204
205
206
207
208
    """The requests that are scheduled from a swap queue.

    Could contain prefill (prefill that's chunked) or decodes.
    """
    # Selected sequences that are going to be swapped in and is in a
    # decoding phase.
    decode_seq_groups: List[SequenceGroup]
    # Selected sequences that are going to be swapped in and in a prefill
    # phase. I.e., it means the prefill has been chunked.
    prefill_seq_groups: List[SequenceGroup]
209
    # The blocks to swap in.
210
    blocks_to_swap_in: List[Tuple[int, int]]
211
    # The blocks to copy.
212
    blocks_to_copy: List[Tuple[int, int]]
213
    # The number of slots for lookahead decoding.
214
    num_lookahead_slots: int
215
216
    # Infeasible sequence groups.
    infeasible_seq_groups: List[SequenceGroup]
217
218
219
220

    @classmethod
    def create_empty(cls) -> "SchedulerSwappedInOutputs":
        return SchedulerSwappedInOutputs(
221
222
            decode_seq_groups=[],
            prefill_seq_groups=[],
223
            blocks_to_swap_in=[],
224
            blocks_to_copy=[],
225
            num_lookahead_slots=0,
226
            infeasible_seq_groups=[],
227
228
229
230
231
        )


@dataclass
class SchedulerPrefillOutputs:
232
233
234
235
236
237
    """The requests that are scheduled from a waiting queue.

    Could contain a fresh prefill requests or preempted requests that need
    to be recomputed from scratch.
    """
    # Selected sequences for prefill.
238
239
240
241
242
243
244
245
246
247
248
249
250
251
    seq_groups: List[SequenceGroup]
    # Ignored sequence groups.
    ignored_seq_groups: List[SequenceGroup]
    num_lookahead_slots: int

    @classmethod
    def create_empty(cls) -> "SchedulerPrefillOutputs":
        return SchedulerPrefillOutputs(
            seq_groups=[],
            ignored_seq_groups=[],
            num_lookahead_slots=0,
        )


Woosuk Kwon's avatar
Woosuk Kwon committed
252
253
class Scheduler:

Woosuk Kwon's avatar
Woosuk Kwon committed
254
    def __init__(
Woosuk Kwon's avatar
Woosuk Kwon committed
255
        self,
256
257
        scheduler_config: SchedulerConfig,
        cache_config: CacheConfig,
258
        lora_config: Optional[LoRAConfig],
Woosuk Kwon's avatar
Woosuk Kwon committed
259
    ) -> None:
260
261
        self.scheduler_config = scheduler_config
        self.cache_config = cache_config
262
263
264
265
        # Note for LoRA scheduling: the current policy is extremely
        # simple and NOT fair. It can lead to starvation of some
        # LoRAs. This should be improved in the future.
        self.lora_config = lora_config
Woosuk Kwon's avatar
Woosuk Kwon committed
266

267
268
269
270
271
272
        if self.scheduler_config.chunked_prefill_enabled:
            self.prompt_limit = self.scheduler_config.max_model_len
        else:
            self.prompt_limit = min(
                self.scheduler_config.max_model_len,
                self.scheduler_config.max_num_batched_tokens)
273

274
275
276
277
278
279
        version = "v1"
        if self.scheduler_config.use_v2_block_manager:
            version = "v2"
        if self.scheduler_config.embedding_mode:
            version = "embedding"

280
        BlockSpaceManagerImpl = BlockSpaceManager.get_block_space_manager_class(
281
            version)
282

Woosuk Kwon's avatar
Woosuk Kwon committed
283
        # Create the block space manager.
284
        self.block_manager = BlockSpaceManagerImpl(
285
286
287
            block_size=self.cache_config.block_size,
            num_gpu_blocks=self.cache_config.num_gpu_blocks,
            num_cpu_blocks=self.cache_config.num_cpu_blocks,
288
289
            sliding_window=self.cache_config.sliding_window,
            enable_caching=self.cache_config.enable_prefix_caching)
290

291
        # Sequence groups in the WAITING state.
292
        # Contain new prefill or preempted requests.
293
        self.waiting: Deque[SequenceGroup] = deque()
294
        # Sequence groups in the RUNNING state.
295
        # Contain decode requests.
296
        self.running: Deque[SequenceGroup] = deque()
297
        # Sequence groups in the SWAPPED state.
298
        # Contain decode requests that are swapped out.
299
        self.swapped: Deque[SequenceGroup] = deque()
Woosuk Kwon's avatar
Woosuk Kwon committed
300

301
302
303
304
305
306
307
        # Time at previous scheduling step
        self.prev_time = 0.0
        # Did we schedule a prompt at previous step?
        self.prev_prompt = False
        # Latency of the last prompt step
        self.last_prompt_latency = 0.0

308
309
310
311
312
313
        # The following field is test-only. It is used to inject artificial
        # preemption.
        self.enable_artificial_preemption = ENABLE_ARTIFICIAL_PREEMPT
        self.artificial_preempt_cnt = (ARTIFICIAL_PREEMPTION_MAX_CNT
                                       if self.enable_artificial_preemption
                                       else 0)
314
        self.num_cumulative_preemption: int = 0
315

316
317
318
319
    @property
    def lora_enabled(self) -> bool:
        return bool(self.lora_config)

320
321
322
323
324
    @property
    def num_decoding_tokens_per_seq(self) -> int:
        """The number of new tokens."""
        return 1

325
    def add_seq_group(self, seq_group: SequenceGroup) -> None:
326
        # Add sequence groups to the waiting queue.
327
        self.waiting.append(seq_group)
Woosuk Kwon's avatar
Woosuk Kwon committed
328

Antoni Baum's avatar
Antoni Baum committed
329
    def abort_seq_group(self, request_id: Union[str, Iterable[str]]) -> None:
330
331
332
333
334
335
336
337
338
339
340
341
        """Aborts a sequence group with the given ID.

        Check if the sequence group with the given ID
            is present in any of the state queue.
        If present, remove the sequence group from the state queue.
            Also, if any of the sequences in the sequence group is not finished,
                free the sequence with status `FINISHED_ABORTED`.
        Otherwise, do nothing.

        Args:
            request_id: The ID(s) of the sequence group to abort.
        """
Antoni Baum's avatar
Antoni Baum committed
342
343
344
        if isinstance(request_id, str):
            request_id = (request_id, )
        request_ids = set(request_id)
345
        for state_queue in [self.waiting, self.running, self.swapped]:
ljss's avatar
ljss committed
346
            aborted_groups: List[SequenceGroup] = []
347
348
349
            for seq_group in state_queue:
                if not request_ids:
                    # Using 'break' here may add two extra iterations,
350
                    # but is acceptable to reduce complexity.
351
                    break
Antoni Baum's avatar
Antoni Baum committed
352
                if seq_group.request_id in request_ids:
353
354
                    # Appending aborted group into pending list.
                    aborted_groups.append(seq_group)
Antoni Baum's avatar
Antoni Baum committed
355
                    request_ids.remove(seq_group.request_id)
356
357
358
            for aborted_group in aborted_groups:
                # Remove the sequence group from the state queue.
                state_queue.remove(aborted_group)
ljss's avatar
ljss committed
359
                for seq in aborted_group.get_seqs():
360
361
362
363
                    if seq.is_finished():
                        continue
                    seq.status = SequenceStatus.FINISHED_ABORTED
                    self.free_seq(seq)
364

365
    def has_unfinished_seqs(self) -> bool:
366
367
        return len(self.waiting) != 0 or len(self.running) != 0 or len(
            self.swapped) != 0
368

369
370
371
    def get_num_unfinished_seq_groups(self) -> int:
        return len(self.waiting) + len(self.running) + len(self.swapped)

372
    def _schedule_running(
373
374
375
376
377
        self,
        running_queue: deque,
        budget: SchedulingBudget,
        curr_loras: Optional[Set[int]],
        policy: Policy,
378
379
380
        enable_chunking: bool = False,
    ) -> Tuple[deque, SchedulerRunningOutputs]:
        """Schedule sequence groups that are running.
381

382
        Running queue should include decode and chunked prefill requests.
Woosuk Kwon's avatar
Woosuk Kwon committed
383

384
385
386
387
388
389
390
391
        Args:
            running_queue: The queue that contains running requests (i.e.,
                decodes). The given arguments are NOT in-place modified.
            budget: The scheduling budget. The argument is in-place updated
                when any decodes are preempted.
            curr_loras: Currently batched lora request ids. The argument is
                in-place updated when any decodes are preempted.
            policy: The sorting policy to sort running_queue.
392
393
394
395
396
            enable_chunking: If True, seq group can be chunked and only a
                chunked number of tokens are scheduled  if
                `budget.num_batched_tokens` has not enough capacity to schedule
                all tokens.
    
397
398
        Returns:
            A tuple of remaining running queue (should be always 0) after
399
            scheduling and SchedulerRunningOutputs.
400
401
        """
        # Blocks that need to be swapped or copied before model execution.
402
        blocks_to_swap_out: List[Tuple[int, int]] = []
403
        blocks_to_copy: List[Tuple[int, int]] = []
Woosuk Kwon's avatar
Woosuk Kwon committed
404

405
406
        decode_seq_groups: List[ScheduledSequenceGroup] = []
        prefill_seq_groups: List[ScheduledSequenceGroup] = []
407
408
        preempted: List[SequenceGroup] = []
        swapped_out: List[SequenceGroup] = []
Woosuk Kwon's avatar
Woosuk Kwon committed
409
410
411

        # NOTE(woosuk): Preemption happens only when there is no available slot
        # to keep all the sequence groups in the RUNNING state.
412
413
        # In this case, the policy is responsible for deciding which sequence
        # groups to preempt.
414
415
416
417
        now = time.time()
        running_queue = policy.sort_by_priority(now, running_queue)
        while running_queue:
            seq_group = running_queue[0]
418
419
420
            num_running_tokens = self._get_num_new_tokens(
                seq_group, SequenceStatus.RUNNING, enable_chunking, budget)

421
422
            if num_running_tokens == 0:
                break
423
424

            running_queue.popleft()
425
            while not self._can_append_slots(seq_group):
426
427
                budget.subtract_num_batched_tokens(seq_group.request_id,
                                                   num_running_tokens)
428
                num_running_seqs = seq_group.get_max_num_running_seqs()
429
430
                budget.subtract_num_seqs(seq_group.request_id,
                                         num_running_seqs)
431
                if curr_loras is not None and seq_group.lora_int_id > 0:
432
                    curr_loras.remove(seq_group.lora_int_id)
433
434

                if running_queue:
435
                    # Preempt the lowest-priority sequence groups.
436
437
438
439
440
441
442
                    victim_seq_group = running_queue.pop()
                    preempted_mode = self._preempt(victim_seq_group,
                                                   blocks_to_swap_out)
                    if preempted_mode == PreemptionMode.RECOMPUTE:
                        preempted.append(victim_seq_group)
                    else:
                        swapped_out.append(victim_seq_group)
443
444
445
                else:
                    # No other sequence groups can be preempted.
                    # Preempt the current sequence group.
446
447
448
449
450
451
                    preempted_mode = self._preempt(seq_group,
                                                   blocks_to_swap_out)
                    if preempted_mode == PreemptionMode.RECOMPUTE:
                        preempted.append(seq_group)
                    else:
                        swapped_out.append(seq_group)
Woosuk Kwon's avatar
Woosuk Kwon committed
452
453
                    break
            else:
454
                self._append_slots(seq_group, blocks_to_copy)
455
456
457
458
459
460
461
462
463
464
465
466
                is_prefill = seq_group.is_prefill()
                if is_prefill:
                    prefill_seq_groups.append(
                        ScheduledSequenceGroup(
                            seq_group=seq_group,
                            token_chunk_size=num_running_tokens))
                else:
                    decode_seq_groups.append(
                        ScheduledSequenceGroup(seq_group=seq_group,
                                               token_chunk_size=1))
                budget.add_num_batched_tokens(seq_group.request_id,
                                              num_running_tokens)
467
468
469
470
471
472
473
                # OPTIMIZATION:  Note that get_max_num_running_seqs is
                # expensive. For the default scheduling chase where
                # enable_chunking is False, num_seqs are updated before running
                # this method, so we don't have to update it again here.
                if enable_chunking:
                    num_running_seqs = seq_group.get_max_num_running_seqs()
                    budget.add_num_seqs(seq_group.request_id, num_running_seqs)
474
475
476
477
478
479
                if curr_loras is not None and seq_group.lora_int_id > 0:
                    curr_loras.add(seq_group.lora_int_id)

        return running_queue, SchedulerRunningOutputs(
            decode_seq_groups=decode_seq_groups,
            prefill_seq_groups=prefill_seq_groups,
480
481
482
483
484
485
            preempted=preempted,
            swapped_out=swapped_out,
            blocks_to_swap_out=blocks_to_swap_out,
            blocks_to_copy=blocks_to_copy,
            num_lookahead_slots=self._get_num_lookahead_slots(
                is_prefill=False))
486

487
488
489
490
491
492
    def _schedule_swapped(
        self,
        swapped_queue: deque,
        budget: SchedulingBudget,
        curr_loras: Optional[Set[int]],
        policy: Policy,
493
        enable_chunking: bool = False,
494
495
    ) -> Tuple[deque, SchedulerSwappedInOutputs]:
        """Schedule sequence groups that are swapped out.
496

497
498
499
        It schedules swapped requests as long as it fits `budget` and
        curr_loras <= max_lora from the scheduling config. The input arguments
        `budget` and `curr_loras` are updated based on scheduled seq_groups.
500

501
502
503
504
505
506
507
508
        Args:
            swapped_queue: The queue that contains swapped out requests.
                The given arguments are NOT in-place modified.
            budget: The scheduling budget. The argument is in-place updated
                when any requests are swapped in.
            curr_loras: Currently batched lora request ids. The argument is
                in-place updated when any requests are swapped in.
            policy: The sorting policy to sort swapped_queue.
509
510
511
512
513
            enable_chunking: If True, seq group can be chunked and only a
                chunked number of tokens are scheduled  if
                `budget.num_batched_tokens` has not enough capacity to schedule
                all tokens.

514
515
516
517
518
        Returns:
            A tuple of remaining swapped_queue after scheduling and
            SchedulerSwappedInOutputs.
        """
        # Blocks that need to be swapped or copied before model execution.
519
        blocks_to_swap_in: List[Tuple[int, int]] = []
520
        blocks_to_copy: List[Tuple[int, int]] = []
521
522
        decode_seq_groups: List[ScheduledSequenceGroup] = []
        prefill_seq_groups: List[ScheduledSequenceGroup] = []
523
524
        now = time.time()
        swapped_queue = policy.sort_by_priority(now, swapped_queue)
525
        infeasible_seq_groups: List[SequenceGroup] = []
526

527
        leftover_swapped: Deque[SequenceGroup] = deque()
528
529
530
531
        while swapped_queue:
            seq_group = swapped_queue[0]

            # If the sequence group cannot be swapped in, stop.
532
533
            alloc_status = self.block_manager.can_swap_in(seq_group)
            if alloc_status == AllocStatus.LATER:
534
                break
535
536
537
538
539
540
541
542
543
544
            elif alloc_status == AllocStatus.NEVER:
                logger.warning(
                    "Failing the request %s because there's not enough kv "
                    "cache blocks to run the entire sequence.",
                    seq_group.request_id)
                for seq in seq_group.get_seqs():
                    seq.status = SequenceStatus.FINISHED_IGNORED
                infeasible_seq_groups.append(seq_group)
                swapped_queue.popleft()
                continue
545
546
547
548

            lora_int_id = 0
            if self.lora_enabled:
                lora_int_id = seq_group.lora_int_id
549
550
551
                assert curr_loras is not None
                assert self.lora_config is not None
                if (lora_int_id > 0 and (lora_int_id not in curr_loras)
552
553
554
555
556
557
558
559
560
561
                        and len(curr_loras) >= self.lora_config.max_loras):
                    # We don't have a space for another LoRA, so
                    # we ignore this request for now.
                    leftover_swapped.appendleft(seq_group)
                    swapped_queue.popleft()
                    continue

            # The total number of sequences in the RUNNING state should not
            # exceed the maximum number of sequences.
            num_new_seqs = seq_group.get_max_num_running_seqs()
562
563
564
            num_new_tokens = self._get_num_new_tokens(seq_group,
                                                      SequenceStatus.SWAPPED,
                                                      enable_chunking, budget)
565

566
567
568
            if (num_new_tokens == 0
                    or not budget.can_schedule(num_new_tokens=num_new_tokens,
                                               num_new_seqs=num_new_seqs)):
569
570
571
572
573
574
575
                break

            if lora_int_id > 0 and curr_loras is not None:
                curr_loras.add(lora_int_id)
            swapped_queue.popleft()
            self._swap_in(seq_group, blocks_to_swap_in)
            self._append_slots(seq_group, blocks_to_copy)
576
577
578
579
580
581
582
583
584
585
            is_prefill = seq_group.is_prefill()
            if is_prefill:
                prefill_seq_groups.append(
                    ScheduledSequenceGroup(seq_group,
                                           token_chunk_size=num_new_tokens))
            else:
                decode_seq_groups.append(
                    ScheduledSequenceGroup(seq_group, token_chunk_size=1))
            budget.add_num_batched_tokens(seq_group.request_id, num_new_tokens)
            budget.add_num_seqs(seq_group.request_id, num_new_seqs)
586
587
588
589

        swapped_queue.extendleft(leftover_swapped)

        return swapped_queue, SchedulerSwappedInOutputs(
590
591
            decode_seq_groups=decode_seq_groups,
            prefill_seq_groups=prefill_seq_groups,
592
593
            blocks_to_swap_in=blocks_to_swap_in,
            blocks_to_copy=blocks_to_copy,
594
            num_lookahead_slots=self._get_num_lookahead_slots(
595
596
597
                is_prefill=False),
            infeasible_seq_groups=infeasible_seq_groups,
        )
598
599
600
601
602
603

    def _schedule_prefills(
        self,
        waiting_queue: deque,
        budget: SchedulingBudget,
        curr_loras: Optional[Set[int]],
604
        enable_chunking: bool = False,
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
    ) -> Tuple[deque, SchedulerPrefillOutputs]:
        """Schedule sequence groups that are in prefill stage.

        Note that the current scheduler treats PREEMPTED_FOR_RECOMPUTE
        as a new prefill (that starts from beginning -> most recently generated
        tokens).

        It schedules waiting requests as long as it fits `budget` and
        curr_loras <= max_lora from the scheduling config. The input arguments
        `budget` and `curr_loras` are updated based on scheduled seq_groups.

        Args:
            waiting_queue: The queue that contains prefill requests.
                The given arguments are NOT in-place modified.
            budget: The scheduling budget. The argument is in-place updated
                when any requests are scheduled.
            curr_loras: Currently batched lora request ids. The argument is
                in-place updated when any requests are scheduled.
623
624
625
626
            enable_chunking: If True, seq group can be chunked and only a
                chunked number of tokens are scheduled  if
                `budget.num_batched_tokens` has not enough capacity to schedule
                all tokens.
627
628
629
630
631
632
633
634
635
636
637

        Returns:
            A tuple of remaining waiting_queue after scheduling and
            SchedulerSwappedInOutputs.
        """
        ignored_seq_groups: List[SequenceGroup] = []
        seq_groups: List[SequenceGroup] = []
        # We don't sort waiting queue because we assume it is sorted.
        # Copy the queue so that the input queue is not modified.
        waiting_queue = deque([s for s in waiting_queue])

638
        leftover_waiting_sequences: Deque[SequenceGroup] = deque()
639
640
641
642
643
644
645
        while self._passed_delay(time.time()) and waiting_queue:
            seq_group = waiting_queue[0]

            waiting_seqs = seq_group.get_seqs(status=SequenceStatus.WAITING)
            assert len(waiting_seqs) == 1, (
                "Waiting sequence group should have only one prompt "
                "sequence.")
646
647
648
649
650
651
652
653
            num_new_tokens = self._get_num_new_tokens(seq_group,
                                                      SequenceStatus.WAITING,
                                                      enable_chunking, budget)
            if not enable_chunking:
                num_prompt_tokens = waiting_seqs[0].get_len()
                assert num_new_tokens == num_prompt_tokens

            if num_new_tokens > self.prompt_limit:
654
                logger.warning(
655
656
657
                    "Input prompt (%d tokens) is too long"
                    " and exceeds limit of %d", num_new_tokens,
                    self.prompt_limit)
658
659
660
661
662
663
664
665
666
667
668
669
                for seq in waiting_seqs:
                    seq.status = SequenceStatus.FINISHED_IGNORED
                ignored_seq_groups.append(seq_group)
                waiting_queue.popleft()
                continue

            # If the sequence group cannot be allocated, stop.
            can_allocate = self.block_manager.can_allocate(seq_group)
            if can_allocate == AllocStatus.LATER:
                break
            elif can_allocate == AllocStatus.NEVER:
                logger.warning(
670
671
672
                    "Input prompt (%d tokens) is too long"
                    " and exceeds the capacity of block_manager",
                    num_new_tokens)
673
674
675
676
677
678
679
680
681
                for seq in waiting_seqs:
                    seq.status = SequenceStatus.FINISHED_IGNORED
                ignored_seq_groups.append(seq_group)
                waiting_queue.popleft()
                continue

            lora_int_id = 0
            if self.lora_enabled:
                lora_int_id = seq_group.lora_int_id
682
683
                assert curr_loras is not None
                assert self.lora_config is not None
684
685
686
687
688
689
690
691
692
693
                if (self.lora_enabled and lora_int_id > 0
                        and lora_int_id not in curr_loras
                        and len(curr_loras) >= self.lora_config.max_loras):
                    # We don't have a space for another LoRA, so
                    # we ignore this request for now.
                    leftover_waiting_sequences.appendleft(seq_group)
                    waiting_queue.popleft()
                    continue

            num_new_seqs = seq_group.get_max_num_running_seqs()
694
695
696
            if (num_new_tokens == 0
                    or not budget.can_schedule(num_new_tokens=num_new_tokens,
                                               num_new_seqs=num_new_seqs)):
697
698
699
700
701
702
                break

            # Can schedule this request.
            if curr_loras is not None and lora_int_id > 0:
                curr_loras.add(lora_int_id)
            waiting_queue.popleft()
703
            self._allocate_and_set_running(seq_group)
704
705
            seq_groups.append(
                ScheduledSequenceGroup(seq_group=seq_group,
706
707
708
                                       token_chunk_size=num_new_tokens))
            budget.add_num_batched_tokens(seq_group.request_id, num_new_tokens)
            budget.add_num_seqs(seq_group.request_id, num_new_seqs)
709
710
711
712
713
714
715
716
717
718
719

        # Queue requests that couldn't be scheduled.
        waiting_queue.extendleft(leftover_waiting_sequences)
        if len(seq_groups) > 0:
            self.prev_prompt = True

        return waiting_queue, SchedulerPrefillOutputs(
            seq_groups=seq_groups,
            ignored_seq_groups=ignored_seq_groups,
            num_lookahead_slots=self._get_num_lookahead_slots(is_prefill=True))

720
721
    def _schedule_default(self) -> SchedulerOutputs:
        """Schedule queued requests.
722
        
723
        The current policy is designed to optimize the throughput. First,
724
725
726
727
728
729
730
731
732
        it batches as many prefill requests as possible. And it schedules
        decodes. If there's a pressure on GPU memory, decode requests can
        be swapped or preempted.
        """
        # Include running requests to the budget.
        budget = SchedulingBudget(
            token_budget=self.scheduler_config.max_num_batched_tokens,
            max_num_seqs=self.scheduler_config.max_num_seqs,
        )
733
734
735
736
737
        # Make sure we include num running seqs before scheduling prefill,
        # so that we don't schedule beyond max_num_seqs for prefill.
        for seq_group in self.running:
            budget.add_num_seqs(seq_group.request_id,
                                seq_group.get_max_num_running_seqs())
738
739
740
741
742
743
        curr_loras = set(
            seq_group.lora_int_id
            for seq_group in self.running) if self.lora_enabled else None

        remaining_waiting, prefills = (self.waiting,
                                       SchedulerPrefillOutputs.create_empty())
744
745
        remaining_running, running_scheduled = (
            self.running, SchedulerRunningOutputs.create_empty())
746
747
748
749
750
751
        remaining_swapped, swapped_in = (
            self.swapped, SchedulerSwappedInOutputs.create_empty())

        # If any requests are swapped, prioritized swapped requests.
        if not self.swapped:
            remaining_waiting, prefills = self._schedule_prefills(
752
                self.waiting, budget, curr_loras, enable_chunking=False)
753

754
        fcfs_policy = PolicyFactory.get_policy(policy_name="fcfs")
755
        # Don't schedule decodes if prefills are scheduled.
756
757
        # NOTE: If `_schedule_prefills` doesn't enable chunking, self.running
        # only contains decode requests, not chunked prefills.
758
        if len(prefills.seq_groups) == 0:
759
760
761
762
763
764
765
            remaining_running, running_scheduled = self._schedule_running(
                self.running,
                budget,
                curr_loras,
                fcfs_policy,
                enable_chunking=False)

766
767
            # If any sequence group is preempted, do not swap in any sequence
            # group. because it means there's no slot for new running requests.
768
769
            if len(running_scheduled.preempted) + len(
                    running_scheduled.swapped_out) == 0:
770
                remaining_swapped, swapped_in = self._schedule_swapped(
771
                    self.swapped, budget, curr_loras, fcfs_policy)
772
773
774
775
776
777
778

        assert (budget.num_batched_tokens <=
                self.scheduler_config.max_num_batched_tokens)
        assert budget.num_curr_seqs <= self.scheduler_config.max_num_seqs

        # Update waiting requests.
        self.waiting = remaining_waiting
779
        self.waiting.extendleft(running_scheduled.preempted)
780
781
782
        # Update new running requests.
        self.running = remaining_running
        self.running.extend([s.seq_group for s in prefills.seq_groups])
783
784
785
786
        self.running.extend(
            [s.seq_group for s in running_scheduled.decode_seq_groups])
        self.running.extend(
            [s.seq_group for s in swapped_in.decode_seq_groups])
787
788
        # Update swapped requests.
        self.swapped = remaining_swapped
789
        self.swapped.extend(running_scheduled.swapped_out)
790
791
        preempted = (len(running_scheduled.preempted) +
                     len(running_scheduled.swapped_out))
792

793
794
795
796
        # There should be no prefill from running queue because this policy
        # doesn't allow chunked prefills.
        assert len(running_scheduled.prefill_seq_groups) == 0
        assert len(swapped_in.prefill_seq_groups) == 0
797
        return SchedulerOutputs(
798
799
800
            scheduled_seq_groups=(prefills.seq_groups +
                                  running_scheduled.decode_seq_groups +
                                  swapped_in.decode_seq_groups),
801
802
803
            num_prefill_groups=len(prefills.seq_groups),
            num_batched_tokens=budget.num_batched_tokens,
            blocks_to_swap_in=swapped_in.blocks_to_swap_in,
804
            blocks_to_swap_out=running_scheduled.blocks_to_swap_out,
805
806
            blocks_to_copy=running_scheduled.blocks_to_copy +
            swapped_in.blocks_to_copy,
807
808
            ignored_seq_groups=prefills.ignored_seq_groups +
            swapped_in.infeasible_seq_groups,
809
            num_lookahead_slots=running_scheduled.num_lookahead_slots,
810
            running_queue_size=len(self.running),
811
            preempted=preempted,
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
        )

    def _schedule_chunked_prefill(self):
        """Schedule queued requests.
        
        Chunked prefill allows to chunk prefill requests, batch them together
        with decode requests. This policy 1. schedule as many decoding requests
        as possible. 2. schedule chunked prefill requests that are not
        finished. 3. schedule swapped request. 4. schedule new prefill
        requests.

        The policy can sustain the high GPU utilization because it can put
        prefill and decodes requests to the same batch, while it improves
        inter token latency because decodes requests don't need to blocked
        by prefill requests.
        """
        budget = SchedulingBudget(
            token_budget=self.scheduler_config.max_num_batched_tokens,
            max_num_seqs=self.scheduler_config.max_num_seqs,
        )
832
        curr_loras: Set[int] = set()
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884

        remaining_waiting, prefills = (self.waiting,
                                       SchedulerPrefillOutputs.create_empty())
        remaining_running, running_scheduled = (
            self.running, SchedulerRunningOutputs.create_empty())
        remaining_swapped, swapped_in = (
            self.swapped, SchedulerSwappedInOutputs.create_empty())

        # Decoding should be always scheduled first by fcfs.
        fcfs_policy = PolicyFactory.get_policy(policy_name="fcfs")
        remaining_running, running_scheduled = self._schedule_running(
            self.running,
            budget,
            curr_loras,
            fcfs_policy,
            enable_chunking=True)

        # Schedule swapped out requests.
        # If preemption happens, it means we don't have space for swap-in.
        if len(running_scheduled.preempted) + len(
                running_scheduled.swapped_out) == 0:
            remaining_swapped, swapped_in = self._schedule_swapped(
                self.swapped, budget, curr_loras, fcfs_policy)

        # Schedule new prefills.
        remaining_waiting, prefills = self._schedule_prefills(
            self.waiting, budget, curr_loras, enable_chunking=True)

        assert (budget.num_batched_tokens <=
                self.scheduler_config.max_num_batched_tokens)
        assert budget.num_curr_seqs <= self.scheduler_config.max_num_seqs

        # Update waiting requests.
        self.waiting = remaining_waiting
        self.waiting.extendleft(running_scheduled.preempted)
        # Update new running requests.
        self.running = remaining_running
        self.running.extend([s.seq_group for s in prefills.seq_groups])
        self.running.extend(
            [s.seq_group for s in running_scheduled.decode_seq_groups])
        self.running.extend(
            [s.seq_group for s in running_scheduled.prefill_seq_groups])
        self.running.extend(
            [s.seq_group for s in swapped_in.decode_seq_groups])
        self.running.extend(
            [s.seq_group for s in swapped_in.prefill_seq_groups])
        # Update swapped requests.
        self.swapped = remaining_swapped
        self.swapped.extend(running_scheduled.swapped_out)
        return SchedulerOutputs(
            scheduled_seq_groups=(prefills.seq_groups +
                                  running_scheduled.prefill_seq_groups +
885
886
887
                                  swapped_in.prefill_seq_groups +
                                  running_scheduled.decode_seq_groups +
                                  swapped_in.decode_seq_groups),
888
889
890
891
892
893
            num_prefill_groups=(len(prefills.seq_groups) +
                                len(swapped_in.prefill_seq_groups) +
                                len(running_scheduled.prefill_seq_groups)),
            num_batched_tokens=budget.num_batched_tokens,
            blocks_to_swap_in=swapped_in.blocks_to_swap_in,
            blocks_to_swap_out=running_scheduled.blocks_to_swap_out,
894
895
            blocks_to_copy=running_scheduled.blocks_to_copy +
            swapped_in.blocks_to_copy,
896
            ignored_seq_groups=prefills.ignored_seq_groups,
897
            num_lookahead_slots=running_scheduled.num_lookahead_slots,
898
            running_queue_size=len(self.running),
899
900
            preempted=(len(running_scheduled.preempted) +
                       len(running_scheduled.swapped_out)),
901
        )
Woosuk Kwon's avatar
Woosuk Kwon committed
902

903
904
905
906
907
908
909
    def _schedule(self) -> SchedulerOutputs:
        """Schedule queued requests."""
        if self.scheduler_config.chunked_prefill_enabled:
            return self._schedule_chunked_prefill()
        else:
            return self._schedule_default()

910
911
912
913
    def _can_append_slots(self, seq_group: SequenceGroup) -> bool:
        """Determine whether or not we have enough space in the KV cache to
        continue generation of the sequence group.
        """
914
915
916
917
918
919
920
        # It is True only for testing case to trigger artificial preemption.
        if (self.enable_artificial_preemption
                and random.uniform(0, 1) < ARTIFICIAL_PREEMPTION_PROB
                and self.artificial_preempt_cnt > 0):
            self.artificial_preempt_cnt -= 1
            return False

921
922
923
924
925
926
927
928
        # Appending slots only occurs in decoding.
        is_prefill = False

        return self.block_manager.can_append_slots(
            seq_group=seq_group,
            num_lookahead_slots=self._get_num_lookahead_slots(is_prefill),
        )

Woosuk Kwon's avatar
Woosuk Kwon committed
929
    def schedule(self) -> Tuple[List[SequenceGroupMetadata], SchedulerOutputs]:
930
931
932
        # Schedule sequence groups.
        # This function call changes the internal states of the scheduler
        # such as self.running, self.swapped, and self.waiting.
Woosuk Kwon's avatar
Woosuk Kwon committed
933
        scheduler_outputs = self._schedule()
934
        now = time.time()
935
936

        # Create input data structures.
937
        seq_group_metadata_list: List[SequenceGroupMetadata] = []
938
939
        for i, scheduled_seq_group in enumerate(
                scheduler_outputs.scheduled_seq_groups):
940
941
            seq_group = scheduled_seq_group.seq_group
            token_chunk_size = scheduled_seq_group.token_chunk_size
942
943
            seq_group.maybe_set_first_scheduled_time(now)

944
            # seq_id -> SequenceData
Light Lin's avatar
Light Lin committed
945
            seq_data: Dict[int, SequenceData] = {}
946
            # seq_id -> physical block numbers
947
            block_tables: Dict[int, List[int]] = {}
948

949
            for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
950
                seq_id = seq.seq_id
951
                seq_data[seq_id] = seq.data
952
                block_tables[seq_id] = self.block_manager.get_block_table(seq)
953
                self.block_manager.access_all_blocks_in_seq(seq, now)
954

955
956
957
958
            common_computed_block_nums = (
                self.block_manager.get_common_computed_block_ids(
                    seq_group.get_seqs(status=SequenceStatus.RUNNING)))

959
960
961
962
963
964
965
966
967
968
969
970
971
972
            do_sample = True
            if seq_group.is_prefill():
                seqs = seq_group.get_seqs()
                # Prefill has only 1 sequence.
                assert len(seqs) == 1
                # In the next iteration, all prompt tokens are not computed.
                # It means the prefill is chunked, and we don't need sampling.
                # NOTE: We use get_len instead of get_prompt_len because when
                # a sequence is preempted, prefill includes previous generated
                # output tokens.
                if (token_chunk_size + seqs[0].data.get_num_computed_tokens() <
                        seqs[0].data.get_len()):
                    do_sample = False

973
974
            # It assumes the scheduled_seq_groups is ordered by
            # prefill < decoding.
975
            is_prompt = seq_group.is_prefill()
976
            seq_group_metadata = SequenceGroupMetadata(
977
                request_id=seq_group.request_id,
978
                is_prompt=is_prompt,
979
                seq_data=seq_data,
980
                sampling_params=seq_group.sampling_params,
981
                block_tables=block_tables,
982
                do_sample=do_sample,
983
                pooling_params=seq_group.pooling_params,
984
                token_chunk_size=token_chunk_size,
985
                lora_request=seq_group.lora_request,
986
                computed_block_nums=common_computed_block_nums,
Nick Hill's avatar
Nick Hill committed
987
                state=seq_group.state,
988
989
990
991
992
                # `multi_modal_data` will only be present for the 1st comm
                # between engine and worker.
                # the subsequent comms can still use delta, but
                # `multi_modal_data` will be None.
                multi_modal_data=seq_group.multi_modal_data
993
                if scheduler_outputs.num_prefill_groups > 0 else None,
994
            )
995
            seq_group_metadata_list.append(seq_group_metadata)
996
997
998
999
1000

        # Now that the batch has been created, we can assume all blocks in the
        # batch will have been computed before the next scheduling invocation.
        # This is because the engine assumes that a failure in model execution
        # will crash the vLLM instance / will not retry.
1001
1002
1003
        for scheduled_seq_group in scheduler_outputs.scheduled_seq_groups:
            self.block_manager.mark_blocks_as_computed(
                scheduled_seq_group.seq_group)
1004

Woosuk Kwon's avatar
Woosuk Kwon committed
1005
        return seq_group_metadata_list, scheduler_outputs
1006

1007
1008
    def fork_seq(self, parent_seq: Sequence, child_seq: Sequence) -> None:
        self.block_manager.fork(parent_seq, child_seq)
Woosuk Kwon's avatar
Woosuk Kwon committed
1009

1010
    def free_seq(self, seq: Sequence) -> None:
1011
        """Free a sequence from a block table."""
1012
        self.block_manager.free(seq)
Woosuk Kwon's avatar
Woosuk Kwon committed
1013

1014
    def free_finished_seq_groups(self) -> None:
1015
1016
        self.running = deque(seq_group for seq_group in self.running
                             if not seq_group.is_finished())
Woosuk Kwon's avatar
Woosuk Kwon committed
1017

1018
    def _allocate_and_set_running(self, seq_group: SequenceGroup) -> None:
1019
        self.block_manager.allocate(seq_group)
1020
        for seq in seq_group.get_seqs(status=SequenceStatus.WAITING):
1021
1022
            seq.status = SequenceStatus.RUNNING

1023
    def _append_slots(
1024
1025
        self,
        seq_group: SequenceGroup,
1026
        blocks_to_copy: List[Tuple[int, int]],
1027
    ) -> None:
1028
1029
1030
1031
1032
        """Appends new slots to the sequences in the given sequence group.

        Args:
            seq_group (SequenceGroup): The sequence group containing the
                sequences to append slots to.
1033
1034
1035
1036
1037
            blocks_to_copy (List[Tuple[int, int]]): A list of tuple of two
                ints, the first int is the source block index, and the second
                int is the destination block index. This list is updated with
                the new source and destination block indices for the appended
                slots.
1038
1039
1040
        """
        num_lookahead_slots = self._get_num_lookahead_slots(is_prefill=False)

1041
        for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
1042
            cows = self.block_manager.append_slots(seq, num_lookahead_slots)
1043
            blocks_to_copy.extend(cows)
1044
1045
1046
1047

    def _preempt(
        self,
        seq_group: SequenceGroup,
1048
        blocks_to_swap_out: List[Tuple[int, int]],
1049
        preemption_mode: Optional[PreemptionMode] = None,
1050
    ) -> PreemptionMode:
1051
1052
1053
        # If preemption mode is not specified, we determine the mode as follows:
        # We use recomputation by default since it incurs lower overhead than
        # swapping. However, when the sequence group has multiple sequences
1054
1055
        # (e.g., beam search), recomputation is not currently supported. In
        # such a case, we use swapping instead.
1056
1057
1058
1059
1060
1061
1062
        # FIXME(woosuk): This makes our scheduling policy a bit bizarre.
        # As swapped sequences are prioritized over waiting sequences,
        # sequence groups with multiple sequences are implicitly prioritized
        # over sequence groups with a single sequence.
        # TODO(woosuk): Support recomputation for sequence groups with multiple
        # sequences. This may require a more sophisticated CUDA kernel.
        if preemption_mode is None:
1063
            if seq_group.get_max_num_running_seqs() == 1:
1064
1065
1066
                preemption_mode = PreemptionMode.RECOMPUTE
            else:
                preemption_mode = PreemptionMode.SWAP
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077

        if self.num_cumulative_preemption % 50 == 0:
            logger.warning(
                "Sequence group %s is preempted by %s mode because there is "
                "not enough KV cache space. This can affect the end-to-end "
                "performance. Increase gpu_memory_utilization or "
                "tensor_parallel_size to provide more KV cache memory. "
                "total_num_cumulative_preemption=%d", seq_group.request_id,
                preemption_mode, self.num_cumulative_preemption + 1)
        self.num_cumulative_preemption += 1

1078
1079
1080
1081
1082
        if preemption_mode == PreemptionMode.RECOMPUTE:
            self._preempt_by_recompute(seq_group)
        elif preemption_mode == PreemptionMode.SWAP:
            self._preempt_by_swap(seq_group, blocks_to_swap_out)
        else:
1083
            raise AssertionError("Invalid preemption mode.")
1084
        return preemption_mode
1085
1086
1087
1088
1089
1090
1091
1092
1093

    def _preempt_by_recompute(
        self,
        seq_group: SequenceGroup,
    ) -> None:
        seqs = seq_group.get_seqs(status=SequenceStatus.RUNNING)
        assert len(seqs) == 1
        for seq in seqs:
            seq.status = SequenceStatus.WAITING
1094
1095
            self.free_seq(seq)
            seq.reset_state_for_recompute()
1096
1097
1098
1099

    def _preempt_by_swap(
        self,
        seq_group: SequenceGroup,
1100
        blocks_to_swap_out: List[Tuple[int, int]],
1101
1102
1103
1104
1105
1106
    ) -> None:
        self._swap_out(seq_group, blocks_to_swap_out)

    def _swap_in(
        self,
        seq_group: SequenceGroup,
1107
        blocks_to_swap_in: List[Tuple[int, int]],
1108
1109
    ) -> None:
        mapping = self.block_manager.swap_in(seq_group)
1110
        blocks_to_swap_in.extend(mapping)
1111
1112
1113
1114
1115
1116
        for seq in seq_group.get_seqs(status=SequenceStatus.SWAPPED):
            seq.status = SequenceStatus.RUNNING

    def _swap_out(
        self,
        seq_group: SequenceGroup,
1117
        blocks_to_swap_out: List[Tuple[int, int]],
1118
    ) -> None:
1119
1120
1121
1122
1123
1124
        if not self.block_manager.can_swap_out(seq_group):
            # FIXME(woosuk): Abort the sequence group instead of aborting the
            # entire engine.
            raise RuntimeError(
                "Aborted due to the lack of CPU swap space. Please increase "
                "the swap space to avoid this error.")
1125
        mapping = self.block_manager.swap_out(seq_group)
1126
        blocks_to_swap_out.extend(mapping)
1127
1128
        for seq in seq_group.get_seqs(status=SequenceStatus.RUNNING):
            seq.status = SequenceStatus.SWAPPED
1129

1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
    def _passed_delay(self, now: float) -> bool:
        if self.prev_prompt:
            self.last_prompt_latency = now - self.prev_time
        self.prev_time, self.prev_prompt = now, False
        # Delay scheduling prompts to let waiting queue fill up
        if self.scheduler_config.delay_factor > 0 and self.waiting:
            earliest_arrival_time = min(
                [e.metrics.arrival_time for e in self.waiting])
            passed_delay = (
                (now - earliest_arrival_time) >
                (self.scheduler_config.delay_factor * self.last_prompt_latency)
                or not self.running)
        else:
            passed_delay = True
        return passed_delay
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157

    def _get_num_lookahead_slots(self, is_prefill: bool) -> int:
        """The number of slots to allocate per sequence per step, beyond known
        token ids. Speculative decoding uses these slots to store KV activations
        of tokens which may or may not be accepted.

        Speculative decoding does not yet support prefill, so we do not perform
        lookahead allocation for prefill.
        """
        if is_prefill:
            return 0

        return self.scheduler_config.num_lookahead_slots
1158
1159
1160

    def _get_num_new_tokens(self, seq_group: SequenceGroup,
                            status: SequenceStatus, enable_chunking: bool,
1161
                            budget: SchedulingBudget) -> int:
1162
1163
1164
1165
1166
1167
1168
        """Get the next new tokens to compute for a given sequence group
            that's in a given `status`.

        The API could chunk the number of tokens to compute based on `budget`
        if `enable_chunking` is True. If a sequence group has multiple
        sequences (e.g., running beam search), it means it is in decoding
        phase, so chunking doesn't happen.
1169
1170

        Returns 0 if the new token cannot be computed due to token budget.
1171
1172
1173
1174
1175
        """
        num_new_tokens = 0
        seqs = seq_group.get_seqs(status=status)
        for seq in seqs:
            num_new_tokens += seq.get_num_new_tokens()
1176
        assert num_new_tokens > 0
1177
1178
1179
1180
1181
1182
1183
        # Chunk if a running request cannot fit in.
        # If number of seq > 1, it means it is doing beam search in a
        # decode phase. Do not chunk in that case.
        if enable_chunking and len(seqs) == 1:
            num_new_tokens = min(num_new_tokens,
                                 budget.remaining_token_budget())
        return num_new_tokens