sampler.py 49 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
"""A layer that samples the next tokens from the model's outputs."""
4
import os
5
import itertools
6
from collections.abc import Iterator
7
from dataclasses import dataclass
8
from importlib.util import find_spec
9
from math import inf
10
from typing import Optional, Union
Woosuk Kwon's avatar
Woosuk Kwon committed
11

12
import msgspec
Woosuk Kwon's avatar
Woosuk Kwon committed
13
14
15
import torch
import torch.nn as nn

16
import vllm.envs as envs
17
from vllm.model_executor.layers.utils import apply_penalties
18
from vllm.model_executor.sampling_metadata import (SamplingMetadata,
19
20
21
                                                   SamplingTensors,
                                                   SequenceGroupToSample)
from vllm.sampling_params import SamplingType
22
23
from vllm.sequence import (VLLM_INVALID_TOKEN_ID,
                           CompletionSequenceGroupOutput, Logprob,
24
                           PromptLogprobs, SampleLogprobs, SequenceOutput)
Woosuk Kwon's avatar
Woosuk Kwon committed
25

26
27
28
29
30
31
32
33
34
if envs.VLLM_USE_FLASHINFER_SAMPLER and find_spec("flashinfer"):
    # yapf: disable
    from flashinfer.sampling import (
        top_k_top_p_sampling_from_probs as flashinfer_top_k_top_p_sampling)

    # yapf: enable
else:
    flashinfer_top_k_top_p_sampling = None

35
36
37
38
from vllm.logger import init_logger

logger = init_logger(__name__)

Joe Runde's avatar
Joe Runde committed
39
40
41
42
43
44

def get_sampler() -> torch.nn.Module:
    if envs.VLLM_USE_V1:
        # Lazy import: the v1 package isn't distributed
        from vllm.v1.sample.sampler import Sampler as V1Sampler
        return V1Sampler()
45
    if envs.VLLM_ZERO_OVERHEAD:
lizhigong's avatar
lizhigong committed
46
        from vllm.zero_overhead.sampler import ZeroOverheadSampler
lizhigong's avatar
lizhigong committed
47
        return ZeroOverheadSampler()
Joe Runde's avatar
Joe Runde committed
48
49
50
    return Sampler()


51
# (num_token_ids, num_parent_ids) per sequence group.
52
SampleResultType = list[tuple[list[int], list[int]]]
53

54
55
# Types of temporary data structures used for
# computing sample_result
56
57
58
59
SampleMetadataType = dict[SamplingType, tuple[list[int],
                                              list[SequenceGroupToSample]]]
MultinomialSamplesType = dict[SamplingType, torch.Tensor]
SampleResultsDictType = dict[int, tuple[list[int], list[int]]]
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


# Encapsulates temporary data structures for computing
# sample_result.
#
# * For multi-step scheduling: must be returned
#   by `Sampler.forward()` and used later to compute the pythonized
#   sample_result
#
# * For single-step scheduling: consumed immediately
#   inside `Sampler.forward()` to compute pythonized sample_result.
@dataclass
class SampleResultArgsType:
    sample_metadata: SampleMetadataType
    multinomial_samples: MultinomialSamplesType
    sample_results_dict: SampleResultsDictType
    sampling_metadata: SamplingMetadata
    greedy_samples: Optional[torch.Tensor]


# Union of non-deferred (single-step scheduling)
# vs deferred (multi-step scheduling)
# sample result types
MaybeDeferredSampleResultType = Union[SampleResultType, SampleResultArgsType]

# Abbreviation of the _sample() return type
86
SampleReturnType = tuple[MaybeDeferredSampleResultType, Optional[torch.Tensor]]
87
88
89
90
91
92
93
94
95
96
97
98
99


class SamplerOutput(
        msgspec.Struct,
        omit_defaults=True,  # type: ignore[call-arg]
        array_like=True):  # type: ignore[call-arg]
    """For each sequence group, we generate a list of SequenceOutput object,
    each of which contains one possible candidate for the next token.

    This data structure implements methods, so it can be used like a list, but
    also has optional fields for device tensors.
    """

100
    outputs: list[CompletionSequenceGroupOutput]
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119

    # On-device tensor containing probabilities of each token.
    sampled_token_probs: Optional[torch.Tensor] = None

    # On-device tensor containing the logprobs of each token.
    logprobs: Optional["torch.Tensor"] = None

    # Holds either (1) the pythonized sampler result (single-step scheduling)
    # or (2) what will be arguments for later deferred pythonization of the
    # sampler result (muliti-step scheduling)
    deferred_sample_results_args: Optional[SampleResultArgsType] = None

    # On-device tensor containing the sampled token ids.
    sampled_token_ids: Optional[torch.Tensor] = None
    # CPU tensor containing the sampled token ids. Used during multi-step to
    # return the sampled token ids from last rank to AsyncLLMEngine to be
    # 'broadcasted' to all other PP ranks for next step.
    sampled_token_ids_cpu: Optional[torch.Tensor] = None

120
121
122
123
124
    # On-device tensor containing the sampled token embeddings (embeddings
    # corresponding to the sampled token ids). Used when prompt embeddings are
    # specified in lieu of prompt token ids or text.
    sampled_token_embeds: Optional[torch.Tensor] = None

125
126
127
128
129
130
131
132
133
134
135
136
137
    # Optional last hidden states from the model.
    hidden_states: Optional[torch.Tensor] = None

    # Optional prefill hidden states from the model
    # (used for models like EAGLE).
    prefill_hidden_states: Optional[torch.Tensor] = None

    # Time taken in the forward pass for this across all workers
    model_forward_time: Optional[float] = None

    # Time taken in the model execute function. This will include model forward,
    # block/sync across workers, cpu-gpu sync time and sampling time.
    model_execute_time: Optional[float] = None
zhuwenwen's avatar
zhuwenwen committed
138
139
140
141
142
143
144
145
146
147
    
    # Optional lm_head logits from the model.
    logits: Optional[torch.Tensor] = None

    # tree-style cartesian candidates
    cart_candidates: Optional[torch.Tensor] = None

    # tree-style cartesian candidates
    tree_attn_masks: Optional[torch.Tensor] = None

148

149
    def __getitem__(self, idx: int) -> CompletionSequenceGroupOutput:
150
151
152
153
154
        return self.outputs[idx]

    def __setitem__(self, idx: int, value):
        self.outputs[idx] = value

155
156
157
    def __iter__(self) -> Iterator[CompletionSequenceGroupOutput]:
        return iter(self.outputs)

158
159
160
161
162
163
164
165
166
167
168
169
170
171
    def __len__(self):
        return len(self.outputs)

    def __eq__(self, other: object):
        return isinstance(other,
                          self.__class__) and self.outputs == other.outputs

    def __repr__(self) -> str:
        """Show the shape of a tensor instead of its values to reduce noise.
        """
        sampled_token_probs_repr = ("None" if self.sampled_token_probs is None
                                    else self.sampled_token_probs.shape)
        sampled_token_ids_repr = ("None" if self.sampled_token_ids is None else
                                  self.sampled_token_ids.shape)
172
173
        return (f"SamplerOutput(outputs={self.outputs}, "
                f"sampled_token_probs={sampled_token_probs_repr}, "
174
175
176
                f"sampled_token_ids={sampled_token_ids_repr})",
                f"logits={self.logits}, ",
                f"tree_attn_masks={self.tree_attn_masks})")
177

178

Woosuk Kwon's avatar
Woosuk Kwon committed
179
class Sampler(nn.Module):
180
181
182
183
184
185
    """Samples the next tokens from the model's outputs.

    This layer does the following:
    1. Discard the hidden states that are not used for sampling (i.e., all
        tokens except the final one in each prompt).
    2. Compute the logits for the next tokens.
186
    3. Apply presence, frequency and repetition penalties.
187
188
189
190
191
    4. Apply temperature scaling.
    5. Apply top-p and top-k truncation.
    6. Sample the next tokens.
    Here, each sequence group within the batch can have different sampling
    parameters (e.g., sampling method, temperature, top-p, top-k, etc.).
192
193
194
195
196
197

    The structure of the logits tensor is coupled with the seq_groups in
    sampling_metadata. Typically, each sequence in each seq_group has one row in
    logits for the next token to be sampled; however, for a seq_group with a
    prompt request with the prompt_logprobs sampling parameter, there are rows
    in logits for each token in the input prompt.
198
    """
Woosuk Kwon's avatar
Woosuk Kwon committed
199

200
201
202
203
204
    def __init__(self):
        super().__init__()

        # Whether or not the SamplerOutput should have on-device tensors
        # containing the sampled token ids and probabilities. This is used by
205
        # speculative decoding and when prompt embeddings are specified.
206
        self.include_gpu_probs_tensor = False
207
        self.should_modify_greedy_probs_inplace = False
208
        self.tree_decoding = (os.environ.get('VLLM_TREE_DECODING') == '1')
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
    def _init_sampling_tensors(
        self,
        logits: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ):
        """The goal here is to reuse sampling tensors between similar decode
        runs. This is possible because sampling logic does not change between
        decodes of the same sequences.
        """
        _, vocab_size = logits.shape

        # First free any existing stored sampling tensors.
        # This is necessary because some sampling tensors may
        # have pinned memory.
        self._sampling_tensors = None

        # Initialize new sampling tensors
        (sampling_tensors, do_penalties, do_top_p_top_k,
         do_min_p) = SamplingTensors.from_sampling_metadata(
             sampling_metadata, vocab_size, logits.device, logits.dtype)

        self._sampling_tensors = sampling_tensors
        self._do_penalties = do_penalties
        self._do_top_p_top_k = do_top_p_top_k
        self._do_min_p = do_min_p

Woosuk Kwon's avatar
Woosuk Kwon committed
236
237
    def forward(
        self,
238
        logits: torch.Tensor,
239
        sampling_metadata: SamplingMetadata,
240
    ) -> Optional[SamplerOutput]:
241
        """
242
        Single-step scheduling:
243
244
245
            * Perform GPU-side sampling computation & compute
            GPU-side logprobs tensor
            * Pythonize sampling result & logprobs tensor
246
247

        Multi-step scheduling:
248
249
250
251
252
253
254
255
            * Perform GPU-side sampling computation & compute
            GPU-side logprobs tensor
            * Defer Pythonization of sampling result & logprobs
            tensor
            * Encapsulate arguments required for deferred Pythonization
            in the
            [`SamplerOutput`][vllm.model_executor.layers.sampler.SamplerOutput]
            structure
256

257
258
259
260
        Args:
            logits: (num_tokens, vocab_size).
            sampling_metadata: Metadata for sampling.
        """
261
        assert logits is not None
262
263
        _, vocab_size = logits.shape

264
        # Prepare sampling tensors with pinned memory to avoid blocking.
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
        if not sampling_metadata.reuse_sampling_tensors:
            self._init_sampling_tensors(logits, sampling_metadata)
        elif self._do_penalties:
            # In this case, the sampling tensors logic depends on
            # "output_tokens" of a sequence. As a result, we cannot
            # reuse sampling tensors, since "output_tokens" changes
            # between decode runs.
            self._init_sampling_tensors(logits, sampling_metadata)

        assert self._sampling_tensors is not None
        sampling_tensors = self._sampling_tensors
        do_penalties = self._do_penalties
        do_top_p_top_k = self._do_top_p_top_k
        do_min_p = self._do_min_p

        logits = _apply_min_tokens_penalty(logits, sampling_metadata)
281

282
        # Apply presence and frequency penalties.
283
        if do_penalties:
284
285
286
287
288
            logits = apply_penalties(logits, sampling_tensors.prompt_tokens,
                                     sampling_tensors.output_tokens,
                                     sampling_tensors.presence_penalties,
                                     sampling_tensors.frequency_penalties,
                                     sampling_tensors.repetition_penalties)
289

290
        # Use float32 to apply temperature scaling.
291
        # Use in-place division to avoid creating a new tensor.
292
        logits = logits.to(torch.float)
293
        logits.div_(sampling_tensors.temperatures.unsqueeze(dim=1))
294

295
        if do_top_p_top_k and flashinfer_top_k_top_p_sampling is None:
296
            logits = _apply_top_k_top_p(logits, sampling_tensors.top_ps,
297
298
                                        sampling_tensors.top_ks)

Roy's avatar
Roy committed
299
        if do_min_p:
300
            logits = _apply_min_p(logits, sampling_tensors.min_ps)
Roy's avatar
Roy committed
301

302
303
304
        # We use float32 for probabilities and log probabilities.
        # Compute the probabilities.
        probs = torch.softmax(logits, dim=-1, dtype=torch.float)
Zhuohan Li's avatar
Zhuohan Li committed
305
306
        # Compute the log probabilities.
        logprobs = torch.log_softmax(logits, dim=-1, dtype=torch.float)
307

Woosuk Kwon's avatar
Woosuk Kwon committed
308
        # Sample the next tokens.
309
        maybe_deferred_sample_results, maybe_sampled_tokens_tensor = _sample(
310
311
312
313
314
315
316
317
318
            probs,
            logprobs,
            sampling_metadata,
            sampling_tensors,
            include_gpu_probs_tensor=self.include_gpu_probs_tensor,
            modify_greedy_probs=self._should_modify_greedy_probs_inplace,
        )

        if self.include_gpu_probs_tensor:
319
320
321
            # Since we will defer sampler result Pythonization,
            # preserve GPU-side tensors in support of later
            # deferred pythonization of logprobs
322
            assert maybe_sampled_tokens_tensor is not None
323
            on_device_tensors = (probs, logprobs, maybe_sampled_tokens_tensor)
324
        else:
325
326
            # Since Pythonization has already happened, don't preserve
            # GPU-side tensors.
327
328
            on_device_tensors = None

329
        # Get the logprobs query results.
330
331
332
        prompt_logprobs = None
        sample_logprobs = None
        if not sampling_metadata.skip_sampler_cpu_output:
333
334
335
336
337
            # Pythonize logprobs now (GPU -> CPU); do not defer.
            assert not isinstance(maybe_deferred_sample_results,
                                  SampleResultArgsType)
            prompt_logprobs, sample_logprobs = get_logprobs(
                logprobs, sampling_metadata, maybe_deferred_sample_results)
338
339

        return _build_sampler_output(
340
            maybe_deferred_sample_results,
341
342
343
344
            sampling_metadata,
            prompt_logprobs,
            sample_logprobs,
            on_device_tensors=on_device_tensors,
345
            skip_sampler_cpu_output=sampling_metadata.skip_sampler_cpu_output,
346
            logits=logits if self.tree_decoding else None)
347
348
349
350
351
352
353
354
355
356
357
358
359

    @property
    def _should_modify_greedy_probs_inplace(self) -> bool:
        """Whether or not the sampler should modify the probability distribution
        of greedily-sampled tokens such that multinomial sampling would sample
        the greedily-sampled token.

        In other words, if True then we set the probability of the greedily-
        sampled token to 1.

        This is used by speculative decoding, which requires that the sampling
        method be encoded into the probability distribution.
        """
360
        return self.should_modify_greedy_probs_inplace
361
362


363
364
365
366
def _apply_min_tokens_penalty(
    logits: torch.Tensor,
    sampling_metadata: SamplingMetadata,
) -> torch.Tensor:
367
368
369
    """Apply min_tokens penalty which sets stop tokens to -inf if min_tokens
        have not been generated yet
    """
370
    # list of indices in logits that will be set to -inf
371
    logits_to_penalize: list[tuple[int, int]] = []
372
373
374
375
376
377
378
379
380
381
    logits_applied = 0
    for seq_group in sampling_metadata.seq_groups:
        seq_ids = seq_group.seq_ids
        sampling_params = seq_group.sampling_params

        sample_indices = seq_group.sample_indices
        logits_applied += len(sample_indices) + len(
            seq_group.prompt_logprob_indices)
        if not seq_group.do_sample:
            continue
382

383
        start_idx = sample_indices[0]
384
        min_tokens = sampling_params.min_tokens
385
386
        token_ids_to_penalize = sampling_params.all_stop_token_ids
        if min_tokens > 0 and token_ids_to_penalize:
387
            seqs_to_penalize: list[int] = []
388
            for j, seq_id in enumerate(seq_ids):
389
                seq_data = seq_group.seq_data[seq_id]
390
                if len(seq_data.output_token_ids_array) < min_tokens:
391
                    seqs_to_penalize.append(j)
392
393
394

            if seqs_to_penalize:
                # convert to the index into logits
395
                seqs_to_penalize = [start_idx + j for j in seqs_to_penalize]
396
397
398
399
400
401
402
403
404
                # itertools.product pairs each seq index with every token id
                logits_to_penalize.extend(
                    itertools.product(seqs_to_penalize, token_ids_to_penalize))

    if logits_to_penalize:
        # use zip and * to group indices along each dimension
        # eg. [ (1,2), (1,3), (5,6) ] -> ( (1,1,5), (2,3,6) )
        logits[tuple(zip(*logits_to_penalize))] = -float("inf")

405
    # verifies that no rows in logits were missed unexpectedly
406
    assert logits_applied == logits.shape[0]
407
408
409
    return logits


410
def _apply_top_k_top_p(
411
    logits: torch.Tensor,
412
413
    p: torch.Tensor,
    k: torch.Tensor,
414
) -> torch.Tensor:
415
416
417
418
419
420
421
422
    logits_sort, logits_idx = logits.sort(dim=-1, descending=False)

    # Apply top-k.
    top_k_mask = logits_sort.size(1) - k.to(torch.long)
    # Get all the top_k values.
    top_k_mask = logits_sort.gather(1, top_k_mask.unsqueeze(dim=1))
    top_k_mask = logits_sort < top_k_mask
    logits_sort.masked_fill_(top_k_mask, -float("inf"))
Woosuk Kwon's avatar
Woosuk Kwon committed
423
424

    # Apply top-p.
425
    probs_sort = logits_sort.softmax(dim=-1)
426
427
428
429
430
    probs_sum = probs_sort.cumsum(dim=-1)
    top_p_mask = probs_sum <= 1 - p.unsqueeze(dim=1)
    # at least one
    top_p_mask[:, -1] = False
    logits_sort.masked_fill_(top_p_mask, -float("inf"))
Woosuk Kwon's avatar
Woosuk Kwon committed
431
432

    # Re-sort the probabilities.
433
434
435
    logits = torch.empty_like(logits_sort).scatter_(dim=-1,
                                                    index=logits_idx,
                                                    src=logits_sort)
436
    return logits
437
438


Roy's avatar
Roy committed
439
440
def _apply_min_p(
    logits: torch.Tensor,
441
    min_p: torch.Tensor,
Roy's avatar
Roy committed
442
443
444
445
446
447
448
) -> torch.Tensor:
    """
    Adapted from
    https://github.com/oobabooga/text-generation-webui/blob/3146124ec01f02c8fb1650a6517cf1b60b537aaf/modules/sampler_hijack.py#L16C17-L16C17
    """
    probs = torch.softmax(logits, dim=-1)
    top_probs, _ = probs.max(dim=-1, keepdim=True)
449
    scaled_min_p = min_p.unsqueeze_(dim=1) * top_probs
Roy's avatar
Roy committed
450
    tokens_to_remove = probs < scaled_min_p
451
    logits = logits.masked_fill_(tokens_to_remove, -float("inf"))
Roy's avatar
Roy committed
452
453
454
455

    return logits


456
def _greedy_sample(
457
    selected_seq_groups: list[SequenceGroupToSample],
458
    samples: torch.Tensor,
459
) -> SampleResultType:
460
461
462
463
464
465
466
467
468
469
470
471
    """Run greedy sampling on a given samples.

    Args:
        selected_seq_groups: A list of sequence groups batched.
        samples: (num_selected_samples,) A tensor of samples. The length of
            samples could be smaller than selected_seq_groups if
            seq_group.do_sample is False.
    Returns:
        Tuple of (next_token_ids, parent_ids). The length of returned list is
        same as the length of selected_seq_groups. If the corresponding
        seq_group has do_sample=False, tuple contains ([], [])
    """
472
    samples_lst = samples.tolist()
473
    sample_idx = 0
474
    results: SampleResultType = []
475
    for seq_group in selected_seq_groups:
476
477
478
479
480
        if not seq_group.do_sample:
            results.append(([], []))
            continue

        seq_ids = seq_group.seq_ids
481
482
483
484
        num_parent_seqs = len(seq_ids)
        assert num_parent_seqs == 1, (
            "Greedy sampling should have only one seq.")
        parent_ids = list(range(num_parent_seqs))
485
        next_token_ids = [samples_lst[sample_idx]]
486
487
488
489
490
491
        results.append((next_token_ids, parent_ids))
        sample_idx += num_parent_seqs
    return results


def _random_sample(
492
    selected_seq_groups: list[SequenceGroupToSample],
493
    random_samples: torch.Tensor,
494
) -> SampleResultType:
495
496
497
498
499
500
501
502
503
504
505
506
    """Run random sampling on a given samples.

    Args:
        selected_seq_groups: A list of sequence groups batched.
        random_samples: (num_selected_samples,) A tensor of samples. The
            length of samples could be smaller than selected_seq_groups if
            seq_group.do_sample is False.
    Returns:
        Tuple of (next_token_ids, parent_ids). The length of returned list is
        same as the length of selected_seq_groups. If the corresponding
        seq_group has do_sample=False, tuple contains ([], [])
    """
507
    # Find the maximum n value of the prompt phase requests.
508
    random_samples = random_samples.cpu()
509
    sample_idx = 0
510
    results: SampleResultType = []
511
512
513
514
515
516
517
518
    for seq_group in selected_seq_groups:
        if not seq_group.do_sample:
            results.append(([], []))
            continue

        seq_ids = seq_group.seq_ids
        sampling_params = seq_group.sampling_params
        is_prompt = seq_group.is_prompt
519
520
521
        num_parent_seqs = len(seq_ids)
        if is_prompt:
            # Prompt phase.
522
            parent_ids = [0] * sampling_params.n
523
            next_token_ids = random_samples[
524
                sample_idx, :sampling_params.n].tolist()
525
526
527
528
529
530
531
532
533
534
        else:
            # Generation phase.
            parent_ids = list(range(num_parent_seqs))
            next_token_ids = random_samples[sample_idx:sample_idx +
                                            num_parent_seqs, 0].tolist()
        results.append((next_token_ids, parent_ids))
        sample_idx += num_parent_seqs
    return results


535
536
537
538
539
540
541
542
# torch.multinomial forces a GPU<->CPU sync.
# Therefore, we use an optimized implementation instead.
# Note that we always sample with replacement.
# probs will be modified in place, but this is fine, as we pass
# in a copy already.
def _multinomial(
    probs: torch.Tensor,
    num_samples: int,
543
    seq_groups: Optional[list[SequenceGroupToSample]] = None,
Nick Hill's avatar
Nick Hill committed
544
) -> torch.Tensor:
545
    if num_samples > 1:
546
        probs = probs.repeat_interleave(num_samples, dim=0)
Nick Hill's avatar
Nick Hill committed
547
548
549
550
551
    q = torch.empty_like(probs)
    if seq_groups is None:
        q.exponential_()
    else:
        sample_idx = 0
552
553
        for seq_group in seq_groups:
            seq_ids = seq_group.seq_ids
554
555
556
557
558
            stride = len(seq_ids) * num_samples
            assert seq_group.generator is not None
            q[sample_idx:sample_idx +
              stride].exponential_(generator=seq_group.generator)
            sample_idx += stride
559
560
561
    return probs.div_(q).argmax(dim=1).view(-1, num_samples)


562
563
def _top_k_top_p_multinomial_with_flashinfer(
        probs: torch.Tensor, top_ks: torch.Tensor, top_ps: torch.Tensor,
564
        num_samples: int, seq_groups: Optional[list[SequenceGroupToSample]]):
565
566
567
568
    if num_samples > 1:
        probs = probs.repeat_interleave(num_samples, dim=0)
        top_ks = top_ks.repeat_interleave(num_samples)
        top_ps = top_ps.repeat_interleave(num_samples)
569
    batch_next_token_ids = flashinfer_top_k_top_p_sampling(
570
571
572
573
574
575
576
        probs,
        top_ks,
        top_ps,
    )
    return batch_next_token_ids.view(-1, num_samples)


577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
def get_pythonized_sample_results(
        sample_result_args: SampleResultArgsType) -> SampleResultType:
    '''This function consumes GPU-side sampler results and computes
    Pythonized CPU-side sampler results (GPU -> CPU sync.)

    Single-step scheduling: this function is invoked at sampling-time
    for immediate Pythonization.

    Multi-step scheduling: Pythonization is deferred until after multiple
    GPU-side steps have been completed.

    Args:
      sample_result_args: GPU-side inputs to the Pythonization process

    Returns:
      Pythonized sampler results
    '''

    (
        sample_metadata,
        sampling_metadata,
        greedy_samples,
        multinomial_samples,
        sample_results_dict,
    ) = (
        sample_result_args.sample_metadata,
        sample_result_args.sampling_metadata,
        sample_result_args.greedy_samples,
        sample_result_args.multinomial_samples,
        sample_result_args.sample_results_dict,
    )

    for sampling_type in SamplingType:
        if sampling_type not in sample_metadata:
            continue
        (seq_group_id, seq_groups) = sample_metadata[sampling_type]
        if sampling_type == SamplingType.GREEDY:
            sample_results = _greedy_sample(seq_groups, greedy_samples)
        elif sampling_type in (SamplingType.RANDOM, SamplingType.RANDOM_SEED):
            sample_results = _random_sample(seq_groups,
                                            multinomial_samples[sampling_type])
        sample_results_dict.update(zip(seq_group_id, sample_results))

    return [
        sample_results_dict.get(i, ([], []))
        for i in range(len(sampling_metadata.seq_groups))
    ]


626
def _sample_with_torch(
627
628
    probs: torch.Tensor,
    logprobs: torch.Tensor,
629
    sampling_metadata: SamplingMetadata,
630
    sampling_tensors: SamplingTensors,
631
632
    include_gpu_probs_tensor: bool,
    modify_greedy_probs: bool,
633
634
635
) -> SampleReturnType:
    '''Torch-oriented _sample() implementation.

636
    Single-step scheduling:
637
638
639
640
641
642
643
644
645
    * Perform GPU-side sampling computation
    * Immediately Pythonize sampling result

    Multi-step scheduling:
    * Perform GPU-side sampling computation
    * Defer Pythonization & preserve GPU-side
      tensors required for Pythonization
    '''

646
    categorized_seq_group_ids: dict[SamplingType, list[int]] = {
647
648
649
        t: []
        for t in SamplingType
    }
650
651
    categorized_sample_indices = sampling_metadata.categorized_sample_indices
    for i, seq_group in enumerate(sampling_metadata.seq_groups):
652
        sampling_params = seq_group.sampling_params
653
654
        sampling_type = sampling_params.sampling_type
        categorized_seq_group_ids[sampling_type].append(i)
655

656
657
658
659
    sample_results_dict: SampleResultsDictType = {}
    sample_metadata: SampleMetadataType = {}
    multinomial_samples: MultinomialSamplesType = {}
    greedy_samples: Optional[torch.Tensor] = None
660

661
662
    # Create output tensor for sampled token ids.
    if include_gpu_probs_tensor:
663
664
665
666
        sampled_token_ids_tensor = torch.full((logprobs.shape[0], 1),
                                              VLLM_INVALID_TOKEN_ID,
                                              dtype=torch.long,
                                              device=logprobs.device)
667
668
669
    else:
        sampled_token_ids_tensor = None

670
671
    # Counterintiutively, having two loops here is actually faster.
    # The first loop can run without waiting on GPU<->CPU sync.
672
    for sampling_type in SamplingType:
673
        sample_indices = categorized_sample_indices[sampling_type]
674
        num_tokens = len(sample_indices)
675
676
        if num_tokens == 0:
            continue
677

678
679
680
681
        seq_group_id = categorized_seq_group_ids[sampling_type]
        seq_groups = [sampling_metadata.seq_groups[i] for i in seq_group_id]
        sample_metadata[sampling_type] = (seq_group_id, seq_groups)
        long_sample_indices = sample_indices.long()
682
        if sampling_type == SamplingType.GREEDY:
683
            greedy_samples = torch.argmax(logprobs[long_sample_indices],
684
                                          dim=-1)
lizhigong's avatar
lizhigong committed
685

686
            if sampled_token_ids_tensor is not None:
687
688
689
690
691
692
693
694
695
696
697
698
                # Store sampled tokens in output tensor.
                sampled_token_ids_tensor[
                    long_sample_indices] = greedy_samples.unsqueeze(-1)

            if modify_greedy_probs:
                # If required, modify the probabilities such that sampling from
                # the modified distribution would always sample the argmax
                # token id.
                _modify_greedy_probs_inplace(logprobs, probs,
                                             long_sample_indices,
                                             greedy_samples)

Nick Hill's avatar
Nick Hill committed
699
        elif sampling_type in (SamplingType.RANDOM, SamplingType.RANDOM_SEED):
700
            max_n_in_batch = 1
701
702
703
            for seq_group in seq_groups:
                if seq_group.is_prompt:
                    sampling_params = seq_group.sampling_params
704
                    max_n_in_batch = max(max_n_in_batch, sampling_params.n)
705
706
707
708
            seq_groups_arg = (None if sampling_type == SamplingType.RANDOM else
                              seq_groups)

            if flashinfer_top_k_top_p_sampling is not None:
709
710
711
712
713
714
715
716
                logger.warning("FlashInfer 0.2.3+ does not support "
                               "per-request generators. Falling back to "
                               "PyTorch-native implementation.")

            multinomial_samples[sampling_type] = _multinomial(
                probs[long_sample_indices],
                max_n_in_batch,
                seq_groups=seq_groups_arg)
lizhigong's avatar
lizhigong committed
717

718
            if sampled_token_ids_tensor is not None:
719
                # Store sampled tokens in output tensor.
720
721
                sampled_token_ids_tensor[long_sample_indices] = \
                    multinomial_samples[sampling_type].to(torch.long)
722

723
724
725
        else:
            raise ValueError(f"Unsupported sampling type: {sampling_type}")

726
727
728
729
730
731
732
733
734
    # Encapsulate arguments for computing Pythonized sampler
    # results, whether deferred or otherwise.
    maybe_deferred_args = SampleResultArgsType(
        sampling_metadata=sampling_metadata,
        sample_metadata=sample_metadata,
        multinomial_samples=multinomial_samples,
        greedy_samples=greedy_samples,
        sample_results_dict=sample_results_dict)

735
    if not sampling_metadata.skip_sampler_cpu_output:
736
737
738
739
740
        # GPU<->CPU sync happens here.
        # This also converts the sampler output to a Python object.
        # Return Pythonized sampler result & sampled token ids
        return get_pythonized_sample_results(
            maybe_deferred_args), sampled_token_ids_tensor
741
    else:
742
743
744
745
746
747
        # Defer sampler result Pythonization; return deferred
        # Pythonization args & sampled token ids
        return (
            maybe_deferred_args,
            sampled_token_ids_tensor,
        )
748
749


750
def _sample(
751
752
753
754
755
756
    probs: torch.Tensor,
    logprobs: torch.Tensor,
    sampling_metadata: SamplingMetadata,
    sampling_tensors: SamplingTensors,
    include_gpu_probs_tensor: bool,
    modify_greedy_probs: bool,
757
) -> SampleReturnType:
758
759
760
761
762
763
764
765
766
767
    """
    Args:
        probs: (num_query_tokens_in_batch, num_vocab)
        logprobs: (num_query_tokens_in_batch, num_vocab)
        sampling_metadata: The metadata for a batch for sampling.
        sampling_tensors: Tensors that include sampling related metadata.

    Returns:
        (next_token_ids, parent_seq_ids) for each seq group in a batch.
            If sampling is skipped, it returns ([], [])
768
        sampled_token_ids_tensor: A tensor of sampled token ids.
769
    """
770
771
772
773
    return _sample_with_torch(
        probs,
        logprobs,
        sampling_metadata,
774
        sampling_tensors,
775
776
777
        include_gpu_probs_tensor=include_gpu_probs_tensor,
        modify_greedy_probs=modify_greedy_probs,
    )
778
779


780
def _get_ranks(x: torch.Tensor, indices: torch.Tensor) -> torch.Tensor:
781
782
783
784
785
786
    """
    This function calculates the ranks of the chosen tokens in a logprob tensor.

    Args:
        x (torch.Tensor): 2D logprob tensor of shape (N, M)
                        where N is the no. of tokens and M is the vocab dim.
787
        indices (torch.Tensor): List of chosen token indices.
788
789
790

    Returns:
        torch.Tensor: 1D tensor of shape (N,) where N is the no. of tokens.
791
                    Each element in the returned tensor represents the rank
792
793
                    of the chosen token in the input logprob tensor.
    """
794
795
    vals = x[torch.arange(0, len(x), device=x.device, dtype=indices.dtype),
             indices]
796
797
798
    result = (x > vals[:, None])
    del vals
    return result.sum(1).add_(1)
799
800


801
def get_logprobs(
802
    logprobs: torch.Tensor,
803
    sampling_metadata: SamplingMetadata,
804
    sample_results: SampleResultType,
805
) -> tuple[list[Optional[PromptLogprobs]], list[SampleLogprobs]]:
806
    """Return sample logprobs and prompt logprobs.
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833

    The logic consists of 3 parts.
    - Select indices to compute logprob from, ranks of token ids, and
        the top k token ids from logprobs.
    - Compute prompt logprobs if required.
    - Compute sample logprobs if required.

    Args:
        logprobs: (num_query_tokens_across_batch, num_vocab). Each query token's
            logprob per vocab. Sequence groups' query tokens are batched in a
            single flattened tensor. For example, assuming there are N
            seq groups, it is sorted by prefill tokens for seq_group_1 (if
            prompt logprob is enabled), decode tokens for seq_group_1 (if
            sampling is required), prefill tokens for seq_group_2, ...
        sampling_metadata: The sampling metadata.
        sample_results: (num_seq_groups) The tuple of (next_token_ids,
            parent_ids) for each sequence group. When beam search is enabled,
            sample_results can contain different number of seq_ids from
            sampling_metadata.seq_groups. It is because beam search creates
            2 * BEAM_WIDTH number of samples (whereas there are only up to
            BEAM_WIDTH number of seq_ids).

    Returns:
        A tuple of prompt and sample logprobs per sequence group in a batch.
    """
    # The index of query token to calculate logprobs. It includes both
    # prompt and sample logprob indices.
834
    query_indices: list[int] = []
835
    # The next token ids to get the logprob value from.
836
    next_token_ids: list[int] = []
837
    # The largest requested number of logprobs. We find logprobs as many as the
838
839
840
    # largest num logprobs in this API. If every logprobs is None, it will be
    # set to -1.
    largest_num_logprobs = -1
841
842
843
844
845
846
847
848
849

    # Select indices to compute logprob from, ranks of token ids, and the top
    # k token ids from logprobs.
    for (seq_group, sample_result) in zip(sampling_metadata.seq_groups,
                                          sample_results):
        sampling_params = seq_group.sampling_params

        # Update indices and tokens for prompt logprobs.
        if (seq_group.is_prompt
850
851
852
                and sampling_params.prompt_logprobs is not None):
            largest_num_logprobs = max(largest_num_logprobs,
                                       sampling_params.prompt_logprobs)
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
            next_prompt_tokens = _get_next_prompt_tokens(seq_group)
            query_indices.extend(seq_group.prompt_logprob_indices)
            next_token_ids.extend(next_prompt_tokens)

        # Update indices and next tokenes for sample logprob.
        if seq_group.do_sample:
            token_ids, parent_seq_ids = sample_result
            # NOTE: We cannot directly use sample_indices because
            # sample_indices only contain parent seq_ids of a previous step.
            # The current step may have different number of seq_ids, and
            # we can obtain it from `sample_result[1]`.
            query_idx = seq_group.sample_indices[0]
            query_indices.extend(
                [query_idx + parent_id for parent_id in parent_seq_ids])
            next_token_ids.extend(token_ids)

            if sampling_params.logprobs is not None:
                largest_num_logprobs = max(largest_num_logprobs,
                                           sampling_params.logprobs)

        assert len(next_token_ids) == len(query_indices)

    if len(query_indices) == 0:
876
877
        empty_sampled_logprob: SampleLogprobs = []
        empty_prompt_logprob: Optional[PromptLogprobs] = None
878
879
880
        num_seq_groups = len(sampling_metadata.seq_groups)
        return [empty_prompt_logprob
                ] * num_seq_groups, [empty_sampled_logprob] * num_seq_groups
881

882
883
884
885
886
    selected_logprobs, ranks = None, None
    top_logprobs, top_token_ids = None, None

    # If largest_num_logprobs == -1, i.e. no logprobs are requested, we can
    # skip the whole logprob calculation.
887
    if largest_num_logprobs >= 0:
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
        query_indices_gpu = torch.tensor(query_indices, device=logprobs.device)
        next_token_ids_gpu = torch.tensor(next_token_ids,
                                          device=logprobs.device)

        # (num_selected_query_tokens, num_logprobs). Note that query_indices can
        # contain duplicates if beam search is enabled.
        selected_logprobs = logprobs[[
            query_indices_gpu,
            next_token_ids_gpu,
        ]]
        ranks = _get_ranks(
            logprobs[query_indices_gpu],
            next_token_ids_gpu,
        )
        assert selected_logprobs.shape[0] == ranks.shape[0]

        # We need to compute top k only if there exists logprobs > 0.
        if largest_num_logprobs > 0:
            # Logprobs of topk tokens for a batch of sequence groups.
            # (num_query_tokens_across_batch).
            top_logprobs, top_token_ids = torch.topk(logprobs,
                                                     largest_num_logprobs,
                                                     dim=-1)
            top_logprobs = top_logprobs.to('cpu')
            top_token_ids = top_token_ids.to('cpu')
913

914
915
        selected_logprobs = selected_logprobs.to('cpu')
        ranks = ranks.to('cpu')
916
917

    # Find prompt/sample logprobs.
918
919
    prompt_logprobs_per_seq_group: list[Optional[PromptLogprobs]] = []
    sample_logprobs_per_seq_group: list[SampleLogprobs] = []
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
    top_logprob_idx = 0
    selected_logprobs_idx = 0

    for seq_group, sample_result in zip(sampling_metadata.seq_groups,
                                        sample_results):
        (prompt_logprobs, top_logprob_idx,
         selected_logprobs_idx) = _get_prompt_logprob_if_needed(
             seq_group, selected_logprobs, ranks, top_token_ids, top_logprobs,
             selected_logprobs_idx, top_logprob_idx)
        prompt_logprobs_per_seq_group.append(prompt_logprobs)

        (sampled_logprobs, top_logprob_idx,
         selected_logprobs_idx) = _get_sampled_logprob_if_needed(
             seq_group, sample_result, selected_logprobs, ranks, top_token_ids,
             top_logprobs, selected_logprobs_idx, top_logprob_idx)
        sample_logprobs_per_seq_group.append(sampled_logprobs)

    return prompt_logprobs_per_seq_group, sample_logprobs_per_seq_group


def _get_prompt_logprob_if_needed(
    seq_group: SequenceGroupToSample,
    selected_logprobs: torch.Tensor,
    ranks: torch.Tensor,
    top_token_ids: torch.Tensor,
    top_logprobs: torch.Tensor,
    selected_logprobs_idx: int,
    top_logprob_idx: int,
):
    """Compute the prompt logprob from a sequence group if needed."""
    sampling_params = seq_group.sampling_params
    is_prompt = seq_group.is_prompt

    # Find prompt logprobs
    prompt_logprobs: Optional[PromptLogprobs] = None
955
    if is_prompt and sampling_params.prompt_logprobs is not None:
956
957
958
        prompt_logprobs = []
        num_logprobs = sampling_params.prompt_logprobs
        next_prompt_tokens = _get_next_prompt_tokens(seq_group)
959
960
961
962
963
964
965
966
967
        # Pre-select indexes and create a list. It is faster than calling .item
        # repetitively.
        selected_logprob_items = selected_logprobs[
            selected_logprobs_idx:selected_logprobs_idx +
            len(next_prompt_tokens)].tolist()
        rank_items = ranks[selected_logprobs_idx:selected_logprobs_idx +
                           len(next_prompt_tokens)].tolist()

        for idx, token_id in enumerate(next_prompt_tokens):
968
969
            # Calculate the prompt logprob of the real prompt tokens.
            # {token_id: (logprob, rank_from_vocab)}
970
            prompt_logprobs_dict: dict[int, tuple[float, int]] = {
971
                token_id: (selected_logprob_items[idx], rank_items[idx])
972
            }
973

974
975
            # Add top K prompt logprobs along with its rank.
            if num_logprobs > 0:
976
977
978
979
980
981
982
983
984
985
986
987
                top_ids = top_token_ids[
                    top_logprob_idx, :num_logprobs].tolist()
                top_probs = top_logprobs[
                    top_logprob_idx, :num_logprobs].tolist()
                # Top K is already sorted by rank, so we can use 1 ~
                # num_logprobs + 1 for rank.
                top_ranks = range(1, num_logprobs + 1)
                prompt_logprobs_dict.update({
                    top_id: (top_prob, rank)
                    for top_id, top_prob, rank in zip(top_ids, top_probs,
                                                      top_ranks)
                })
988
989
990
991
992
993
            prompt_logprobs.append({
                token_id: Logprob(*logprob_and_rank)
                for token_id, logprob_and_rank in prompt_logprobs_dict.items()
            })
            # + 1 to go to the next prompt token.
            top_logprob_idx += 1
994
995
996

        # + len(next_prompt_tokens) to go to the next prompt.
        selected_logprobs_idx += len(next_prompt_tokens)
997
998
999
1000
1001
    return prompt_logprobs, top_logprob_idx, selected_logprobs_idx


def _get_sampled_logprob_if_needed(
    seq_group: SequenceGroupToSample,
1002
    sample_result: tuple[list[int], list[int]],
1003
1004
1005
1006
1007
1008
1009
1010
1011
    selected_logprobs: torch.Tensor,
    ranks: torch.Tensor,
    top_token_ids: torch.Tensor,
    top_logprobs: torch.Tensor,
    selected_logprobs_idx: int,
    top_logprob_idx: int,
):
    """Compute the sample logprob if needed."""
    seq_ids = seq_group.seq_ids
1012
    num_logprobs = seq_group.sampling_params.logprobs
1013
1014
1015
1016
1017
    sampled_logprobs: SampleLogprobs = []
    next_token_ids, parent_seq_ids = sample_result

    if seq_group.do_sample:
        assert len(next_token_ids) > 0
1018
        if num_logprobs is None:
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
            for next_token_id in next_token_ids:
                # Use a dummy logprob
                sampled_logprobs.append({next_token_id: Logprob(inf)})
        else:
            # Pre-select items from tensor. tolist() is faster than repetitive
            # `.item()` calls.
            selected_logprob_items = selected_logprobs[
                selected_logprobs_idx:selected_logprobs_idx +
                len(next_token_ids)].tolist()
            rank_items = ranks[selected_logprobs_idx:selected_logprobs_idx +
                               len(next_token_ids)].tolist()
            for idx, (next_token_id, parent_id) in enumerate(
                    zip(next_token_ids, parent_seq_ids)):
                # Get the logprob of a sampled token.
                sampled_logprobs_dict = {
                    next_token_id:
                    (selected_logprob_items[idx], rank_items[idx])
                }
                if num_logprobs is not None and num_logprobs > 0:
                    # Get top K logprobs.
                    top_ids = top_token_ids[top_logprob_idx +
                                            parent_id, :num_logprobs].tolist()
                    top_probs = top_logprobs[
                        top_logprob_idx + parent_id, :num_logprobs].tolist()
                    # Top K is already sorted by rank, so we can use 1 ~
                    # num_logprobs + 1 for rank.
                    top_ranks = range(1, num_logprobs + 1)
                    sampled_logprobs_dict.update({
                        top_id: (top_prob, rank)
                        for top_id, top_prob, rank in zip(
                            top_ids, top_probs, top_ranks)
                    })

                sampled_logprobs.append({
                    token_id: Logprob(*logprob_and_rank)
                    for token_id, logprob_and_rank in
                    sampled_logprobs_dict.items()
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
                })

        # NOTE: This part of code is not intuitive. `selected_logprobs` include
        # logprobs for the current step, which has len(next_token_ids) tokens
        # per sequence group. `logprobs` includes logprobs from the previous
        # steps, which has len(seq_ids) tokens per sequence group.

        # Iterate to the next sequence group in a batch.
        selected_logprobs_idx += len(next_token_ids)
        # Iterate to the next sequence group in a batch.
1066
1067
        top_logprob_idx += len(seq_ids)
    return sampled_logprobs, top_logprob_idx, selected_logprobs_idx
1068
1069


1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
def _modify_greedy_probs_inplace(logprobs: torch.Tensor, probs: torch.Tensor,
                                 sample_indices: torch.Tensor,
                                 greedy_samples: torch.Tensor) -> None:
    """Modify the probability distributions of the greedily-sampled tokens such
    that each sampled token has a "probability" of 1.0. This is required by
    speculative decoding, which depends on the sampling method being encoded
    within the probability distribution for correctness.

    # Why do we only need to do this for greedy sampling?

    vLLM's sampler performs the following steps for greedy or multinomial
    (random) sampling:
        1. Get logits from model.
        2. Modify logits according to per-sequence sampling parameters.
            - Multiply by temperature, top-k and top-p masking, penalize tokens
                according to their frequency, etc.
        3. Sample a token.
            - Random sampling simply samples from the modified probability
                distribution.
            - Greedy sampling performs `argmax` to obtain the token with the
                highest likelihood.
1091

1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
    Ignoring greedy sampling for a moment, we find that the computed probability
    distribution has the following property: we can sample from it independently
    and find that the token sampled by the Sampler has a frequency corresponding
    to how often we see it in our sampling. In other words, for tokens sampled
    with vLLM's random SamplingType, the computed probability distribution
    encodes the sampling methodology completely.

    Greedy sampling does not normally have this property. vLLM modifies logits
    according to sampling params, then performs `argmax`, then returns the
    sampled token and the computed probability distribution. If we sample from
    the distribution, we'll find the likelihood of the greedily-sampled token
    is not always 1.0.

    Since lossless speculative decoding requires that the sampling methodology
    be encoded within the probability distribution, we are motivated to modify
    the probability distribution such that the sampled token has probability 1
    when speculative decoding is used.

    NOTE: Alternatively, we could use an extremely low temperature to achieve
    greedy sampling using multinomial computation and unite the codepaths. This
    has implications on the overall design of the sampler, e.g. how to record
    accurate logprobs for the user, so this improvement is deferred to later.
    """
1115
    # NOTE: logprobs are not modified so they can be returned to the user.
1116
1117
1118
1119
    probs[sample_indices, :] = 0
    probs[sample_indices, greedy_samples] = 1.0


1120
def _build_sampler_output(
1121
    maybe_deferred_sample_results: MaybeDeferredSampleResultType,
1122
    sampling_metadata: SamplingMetadata,
1123
1124
1125
    prompt_logprobs: Optional[list[Optional[PromptLogprobs]]],
    sample_logprobs: Optional[list[SampleLogprobs]],
    on_device_tensors: Optional[tuple[torch.Tensor, torch.Tensor,
1126
                                      torch.Tensor]],
1127
    skip_sampler_cpu_output: bool = False,
1128
    logits: Optional[torch.Tensor] = None
1129
) -> SamplerOutput:
1130
1131
1132
1133
1134
1135
1136
1137
    """Construct Python objects with the output of sampling.

    Args:
        on_device_tensors: Tuple containing on-device tensors with the
            probabilities used in sampling and the sampled token ids. This
            allows post-processing without copies to CPU/serialization, e.g. in
            speculative decoding rejection sampling.
    """
1138
    sampler_output: list[CompletionSequenceGroupOutput] = []
1139
1140
1141
1142
1143

    if skip_sampler_cpu_output:
        assert isinstance(maybe_deferred_sample_results, SampleResultArgsType)
        deferred_sample_results_args = maybe_deferred_sample_results
    else:
1144
1145
        assert prompt_logprobs is not None
        assert sample_logprobs is not None
1146
1147
        assert not isinstance(maybe_deferred_sample_results,
                              SampleResultArgsType)
1148
1149
1150
1151
        assert len(sampling_metadata.seq_groups) \
            == len(maybe_deferred_sample_results) \
            == len(prompt_logprobs) \
            == len(sample_logprobs)
1152
        deferred_sample_results_args = None
1153
1154
1155

        for (seq_group, sample_result, group_prompt_logprobs,
             group_sample_logprobs) in zip(sampling_metadata.seq_groups,
1156
1157
                                           maybe_deferred_sample_results,
                                           prompt_logprobs, sample_logprobs):
1158
1159
            seq_ids = seq_group.seq_ids
            next_token_ids, parent_ids = sample_result
1160
            seq_outputs: list[SequenceOutput] = []
1161
1162
1163
1164
1165
1166
1167
1168
            for parent_id, next_token_id, logprobs in zip(
                    parent_ids, next_token_ids, group_sample_logprobs):
                seq_outputs.append(
                    SequenceOutput(seq_ids[parent_id], next_token_id,
                                   logprobs))
            sampler_output.append(
                CompletionSequenceGroupOutput(seq_outputs,
                                              group_prompt_logprobs))
1169
1170
1171

    # If not specified, store None values in SamplerOutput.
    if on_device_tensors is not None:
1172
1173
        (sampled_token_probs, logprobs_tensor,
         sampled_token_ids) = on_device_tensors
1174
    else:
1175
1176
        sampled_token_probs, logprobs_tensor, sampled_token_ids = (None, None,
                                                                   None)
1177
1178
1179
1180
1181

    return SamplerOutput(
        outputs=sampler_output,
        sampled_token_probs=sampled_token_probs,
        sampled_token_ids=sampled_token_ids,
1182
        logprobs=logprobs_tensor,
1183
1184
        deferred_sample_results_args=deferred_sample_results_args,
        logits=logits)
1185
1186


1187
1188
def _get_next_prompt_tokens(
        seq_group: SequenceGroupToSample) -> tuple[int, ...]:
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
    """Get a list of next prompt tokens to compute logprob from a
        given sequence group.

    It is used to compute prompt logprob. Imagine you have logprob for each
    query token. Query token needs to know the next prompt token id to compute
    prompt logprob. This is a helper to obtain next prompt token ids.

    This API has to be used only when the caller knows seq_group is in prefill
    stage.

    Returns:
        A list of next prompt tokens to compute logprob.
    """
    assert seq_group.is_prompt, (
        "Caller should ensure the sequence group is in a prefill stage.")
    seq_ids = seq_group.seq_ids
1205
1206
    query_len = seq_group.query_len
    assert query_len is not None
1207
1208
1209
1210
1211
1212
1213
    # prompt has only 1 seq id.
    assert len(seq_ids) == 1
    seq_data = seq_group.seq_data[seq_ids[0]]
    computed_len = seq_data.get_num_computed_tokens()
    prompt_tokens = seq_data.prompt_token_ids
    # +1 because we are looking for a next prompt token.
    next_token_index_start = computed_len + 1
1214
    next_token_index_end = min(computed_len + query_len + 1,
1215
1216
1217
1218
                               len(prompt_tokens))
    next_prompt_tokens = prompt_tokens[
        next_token_index_start:next_token_index_end]
    return next_prompt_tokens