sampler.py 43.6 KB
Newer Older
1
"""A layer that samples the next tokens from the model's outputs."""
2
import itertools
3
from typing import Dict, List, Optional, Tuple
Woosuk Kwon's avatar
Woosuk Kwon committed
4
5
6
7

import torch
import torch.nn as nn

8
from vllm.model_executor.layers.ops.sample import sample as sample_triton
9
from vllm.model_executor.sampling_metadata import (SamplingMetadata,
10
11
12
                                                   SamplingTensors,
                                                   SequenceGroupToSample)
from vllm.sampling_params import SamplingType
13
from vllm.sequence import (Logprob, PromptLogprobs, SampleLogprobs,
14
                           SamplerOutput, SequenceGroupOutput, SequenceOutput)
Woosuk Kwon's avatar
Woosuk Kwon committed
15

16

Woosuk Kwon's avatar
Woosuk Kwon committed
17
class Sampler(nn.Module):
18
19
20
21
22
23
    """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.
24
    3. Apply presence, frequency and repetition penalties.
25
26
27
28
29
    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.).
30
31
32
33
34
35

    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.
36
    """
Woosuk Kwon's avatar
Woosuk Kwon committed
37

38
39
40
41
42
43
44
45
    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
        # speculative decoding.
        self.include_gpu_probs_tensor = False

Woosuk Kwon's avatar
Woosuk Kwon committed
46
47
    def forward(
        self,
48
        logits: torch.Tensor,
49
        sampling_metadata: SamplingMetadata,
50
    ) -> Optional[SamplerOutput]:
51
52
53
54
55
        """
        Args:
            logits: (num_tokens, vocab_size).
            sampling_metadata: Metadata for sampling.
        """
56
        assert logits is not None
57
58
        _, vocab_size = logits.shape

59
60
        logits = _apply_min_tokens_penalty(logits, sampling_metadata)

61
62
63
64
        # Prepare sampling tensors with pinned memory to avoid blocking.
        (sampling_tensors, do_penalties, do_top_p_top_k,
         do_min_p) = SamplingTensors.from_sampling_metadata(
             sampling_metadata, vocab_size, logits.device, logits.dtype)
65

66
        # Apply presence and frequency penalties.
67
68
69
70
71
72
        if do_penalties:
            logits = _apply_penalties(logits, sampling_tensors.prompt_tokens,
                                      sampling_tensors.output_tokens,
                                      sampling_tensors.presence_penalties,
                                      sampling_tensors.frequency_penalties,
                                      sampling_tensors.repetition_penalties)
73

74
        # Apply temperature scaling.
75
76
77
78
        # Use in-place division to avoid creating a new tensor.
        logits.div_(sampling_tensors.temperatures.unsqueeze_(dim=1))

        if do_top_p_top_k:
79
            logits = _apply_top_k_top_p(logits, sampling_tensors.top_ps,
80
81
                                        sampling_tensors.top_ks)

Roy's avatar
Roy committed
82
        if do_min_p:
83
            logits = _apply_min_p(logits, sampling_tensors.min_ps)
Roy's avatar
Roy committed
84

85
86
87
        # 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
88
89
        # Compute the log probabilities.
        logprobs = torch.log_softmax(logits, dim=-1, dtype=torch.float)
90

Woosuk Kwon's avatar
Woosuk Kwon committed
91
        # Sample the next tokens.
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
        sample_results, maybe_sampled_tokens_tensor = _sample(
            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:
            assert maybe_sampled_tokens_tensor is not None
            sampled_tokens_tensor = maybe_sampled_tokens_tensor
            on_device_tensors = (probs, sampled_tokens_tensor)
        else:
            on_device_tensors = None

108
109
        # Get the logprobs query results.
        prompt_logprobs, sample_logprobs = _get_logprobs(
110
            logprobs, sampling_metadata, sample_results)
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
        return _build_sampler_output(sample_results,
                                     sampling_metadata,
                                     prompt_logprobs,
                                     sample_logprobs,
                                     on_device_tensors=on_device_tensors)

    @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.
        """
        # Modify greedy probs if include_gpu_probs_tensor is set.
        return self.include_gpu_probs_tensor
131
132


133
def _get_bin_counts_and_mask(
134
    tokens: torch.Tensor,
135
136
137
138
139
140
141
    vocab_size: int,
    num_seqs: int,
) -> Tuple[torch.Tensor, torch.Tensor]:
    # Compute the bin counts for the tokens.
    # vocab_size + 1 for padding.
    bin_counts = torch.zeros((num_seqs, vocab_size + 1),
                             dtype=torch.long,
142
143
                             device=tokens.device)
    bin_counts.scatter_add_(1, tokens, torch.ones_like(tokens))
144
145
146
147
    bin_counts = bin_counts[:, :vocab_size]
    mask = bin_counts > 0

    return bin_counts, mask
148
149


150
151
152
153
def _apply_min_tokens_penalty(
    logits: torch.Tensor,
    sampling_metadata: SamplingMetadata,
) -> torch.Tensor:
154
155
156
    """Apply min_tokens penalty which sets stop tokens to -inf if min_tokens
        have not been generated yet
    """
157
158
    # list of indices in logits that will be set to -inf
    logits_to_penalize = []
159
160
161
162
163
164
165
166
167
168
    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
169

170
        start_idx = sample_indices[0]
171
        min_tokens = sampling_params.min_tokens
172
173
        token_ids_to_penalize = sampling_params.all_stop_token_ids
        if min_tokens > 0 and token_ids_to_penalize:
174
            seqs_to_penalize = []
175
            for j, seq_id in enumerate(seq_ids):
176
                seq_data = seq_group.seq_data[seq_id]
177
                if len(seq_data.output_token_ids) < min_tokens:
178
                    seqs_to_penalize.append(j)
179
180
181

            if seqs_to_penalize:
                # convert to the index into logits
182
                seqs_to_penalize = [start_idx + j for j in seqs_to_penalize]
183
184
185
186
187
188
189
190
191
                # 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")

192
    # verifies that no rows in logits were missed unexpectedly
193
    assert logits_applied == logits.shape[0]
194
195
196
    return logits


197
198
199
200
201
def _apply_penalties(logits: torch.Tensor, prompt_tokens_tensor: torch.Tensor,
                     output_tokens_tensor: torch.Tensor,
                     presence_penalties: torch.Tensor,
                     frequency_penalties: torch.Tensor,
                     repetition_penalties: torch.Tensor) -> torch.Tensor:
202
    num_seqs, vocab_size = logits.shape
203
204
    _, prompt_mask = _get_bin_counts_and_mask(prompt_tokens_tensor, vocab_size,
                                              num_seqs)
205
    output_bin_counts, output_mask = _get_bin_counts_and_mask(
206
        output_tokens_tensor, vocab_size, num_seqs)
207

ljss's avatar
ljss committed
208
    repetition_penalties = repetition_penalties[:, None].repeat(1, vocab_size)
209
    repetition_penalties[~(prompt_mask | output_mask)] = 1.0
ljss's avatar
ljss committed
210
211
212
    logits = torch.where(logits > 0, logits / repetition_penalties,
                         logits * repetition_penalties)

213
214
    # We follow the definition in OpenAI API.
    # Refer to https://platform.openai.com/docs/api-reference/parameter-details
215
216
    logits -= frequency_penalties.unsqueeze_(dim=1) * output_bin_counts
    logits -= presence_penalties.unsqueeze_(dim=1) * output_mask
217
218
219
    return logits


220
def _apply_top_k_top_p(
221
    logits: torch.Tensor,
222
223
    p: torch.Tensor,
    k: torch.Tensor,
224
) -> torch.Tensor:
225
226
227
228
229
230
231
232
    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
233
234

    # Apply top-p.
235
    probs_sort = logits_sort.softmax(dim=-1)
236
237
238
239
240
    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
241
242

    # Re-sort the probabilities.
243
244
245
246
247
248
    src = torch.arange(logits_idx.shape[-1],
                       device=logits_idx.device).expand_as(logits_idx)
    logits_idx_inv = torch.empty_like(logits_idx).scatter_(dim=-1,
                                                           index=logits_idx,
                                                           src=src)
    logits = torch.gather(logits_sort, dim=-1, index=logits_idx_inv)
249
    return logits
250
251


Roy's avatar
Roy committed
252
253
def _apply_min_p(
    logits: torch.Tensor,
254
    min_p: torch.Tensor,
Roy's avatar
Roy committed
255
256
257
258
259
260
261
) -> 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)
262
    scaled_min_p = min_p.unsqueeze_(dim=1) * top_probs
Roy's avatar
Roy committed
263
    tokens_to_remove = probs < scaled_min_p
264
    logits = logits.masked_fill_(tokens_to_remove, -float("inf"))
Roy's avatar
Roy committed
265
266
267
268

    return logits


269
def _greedy_sample(
270
    selected_seq_groups: List[SequenceGroupToSample],
271
    samples: torch.Tensor,
272
) -> List[Tuple[List[int], List[int]]]:
273
274
275
276
277
278
279
280
281
282
283
284
    """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 ([], [])
    """
285
    samples = samples.tolist()
286
287
288
    sample_idx = 0
    results = []
    for seq_group in selected_seq_groups:
289
290
291
292
293
        if not seq_group.do_sample:
            results.append(([], []))
            continue

        seq_ids = seq_group.seq_ids
294
295
296
297
        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))
298
        next_token_ids = [samples[sample_idx]]
299
300
301
302
303
304
        results.append((next_token_ids, parent_ids))
        sample_idx += num_parent_seqs
    return results


def _random_sample(
305
    selected_seq_groups: List[SequenceGroupToSample],
306
    random_samples: torch.Tensor,
307
) -> List[Tuple[List[int], List[int]]]:
308
309
310
311
312
313
314
315
316
317
318
319
    """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 ([], [])
    """
320
    # Find the maximum best_of value of the prompt phase requests.
321
    random_samples = random_samples.cpu()
322
323
    sample_idx = 0
    results = []
324
325
326
327
328
329
330
331
    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
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
        num_parent_seqs = len(seq_ids)
        if is_prompt:
            # Prompt phase.
            parent_ids = [0] * sampling_params.best_of
            next_token_ids = random_samples[
                sample_idx, :sampling_params.best_of].tolist()
        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


def _beam_search_sample(
349
    selected_seq_groups: List[SequenceGroupToSample],
350
    logprobs: torch.Tensor,
351
) -> List[Tuple[List[int], List[int]]]:
352
353
354
355
356
357
358
359
360
361
362
    """Run beam sampling on a given samples.

    Args:
        selected_seq_groups: A list of sequence groups batched.
        logprobs: (num_selected_samples, vocab_size,) A tensor of logprob
        on selected sample indices.
    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 ([], [])
    """
363
364
365
366
367
368
369
    # We sample 2 * beam_width candidates to make sure that with high
    # probability we can get `beam_width` candidates in addition to
    # the finished sequences for the next iteration. See
    # https://github.com/tensorflow/tensor2tensor/blob/bafdc1b67730430d38d6ab802cbd51f9d053ba2e/tensor2tensor/utils/beam_search.py#L557-L563
    # for details. See also HF reference:
    # https://github.com/huggingface/transformers/blob/a4dd53d88e4852f023332d284ff07a01afcd5681/src/transformers/generation/utils.py#L3063-L3065
    #
370
    # NOTE: Beam search is not vectorized, so its speed can be slower than
371
372
373
    # other sampling methods.
    sample_idx = 0
    results = []
374
375
376
377
378
379
380
    for seq_group in selected_seq_groups:
        if not seq_group.do_sample:
            results.append(([], []))
            continue

        is_prompt = seq_group.is_prompt
        seq_ids, sampling_params = seq_group.seq_ids, seq_group.sampling_params
381
382
383
384
385
386
387
388
389
390
391
392
393
394
        num_parent_seqs = len(seq_ids)
        beam_width = sampling_params.best_of
        seq_group_logprobs = logprobs[sample_idx:sample_idx + num_parent_seqs]
        if is_prompt:
            # Prompt phase.
            assert num_parent_seqs == 1, (
                "Prompt input should have only one seq.")
            parent_ids = [0] * (2 * beam_width)
            _, next_token_ids = torch.topk(seq_group_logprobs[0],
                                           2 * beam_width)
            next_token_ids = next_token_ids.tolist()
        else:
            # Generation phase.
            cumulative_logprobs = [
395
396
                seq_group.seq_data[seq_id].cumulative_logprob
                for seq_id in seq_ids
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
            ]
            cumulative_logprobs = torch.tensor(
                cumulative_logprobs,
                dtype=torch.float,
                device=seq_group_logprobs.device)
            seq_group_logprobs = (seq_group_logprobs +
                                  cumulative_logprobs.unsqueeze(dim=1))
            _, topk_ids = torch.topk(seq_group_logprobs.flatten(),
                                     2 * beam_width)
            topk_ids = topk_ids.tolist()
            vocab_size = seq_group_logprobs.size(-1)
            parent_ids = [i // vocab_size for i in topk_ids]
            next_token_ids = [i % vocab_size for i in topk_ids]
        results.append((next_token_ids, parent_ids))
        sample_idx += num_parent_seqs
    assert sample_idx == logprobs.size(0)
    return results
414
415


416
417
418
419
420
421
422
423
# 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,
424
    seq_groups: Optional[List[SequenceGroupToSample]] = None,
Nick Hill's avatar
Nick Hill committed
425
) -> torch.Tensor:
426
427
428
429
430
431
432
433
434
    if num_samples > 1:
        # This is equivalent to torch.repeat_interleaved (which also
        # forces a GPU<->CPU sync).
        # This allows us to do sampling with replacement by creating
        # num_samples copies of each row in the tensor, and then
        # batch sampling the resulting tensor.
        probs = probs[:, None, :].expand(probs.shape[0], num_samples,
                                         probs.shape[1]).contiguous().view(
                                             -1, probs.shape[1])
Nick Hill's avatar
Nick Hill committed
435
436
437
438
439
    q = torch.empty_like(probs)
    if seq_groups is None:
        q.exponential_()
    else:
        sample_idx = 0
440
441
        for seq_group in seq_groups:
            seq_ids = seq_group.seq_ids
Nick Hill's avatar
Nick Hill committed
442
            next_sample_idx = sample_idx + len(seq_ids) * num_samples
443
444
            q[sample_idx:next_sample_idx].exponential_(
                generator=seq_group.generator)
Nick Hill's avatar
Nick Hill committed
445
            sample_idx = next_sample_idx
446
447
448
    return probs.div_(q).argmax(dim=1).view(-1, num_samples)


449
def _sample_with_torch(
450
451
    probs: torch.Tensor,
    logprobs: torch.Tensor,
452
    sampling_metadata: SamplingMetadata,
453
454
455
    include_gpu_probs_tensor: bool,
    modify_greedy_probs: bool,
) -> Tuple[List[Tuple[List[int], List[int]]], Optional[torch.Tensor]]:
456
    categorized_seq_group_ids = {t: [] for t in SamplingType}
457
458
    categorized_sample_indices = sampling_metadata.categorized_sample_indices
    for i, seq_group in enumerate(sampling_metadata.seq_groups):
459
        sampling_params = seq_group.sampling_params
460
461
        sampling_type = sampling_params.sampling_type
        categorized_seq_group_ids[sampling_type].append(i)
462
463

    sample_results_dict: Dict[int, Tuple[List[int], List[int]]] = {}
464
    sample_metadata = {}
Nick Hill's avatar
Nick Hill committed
465
    multinomial_samples = {}
466

467
468
469
470
471
472
473
474
475
    # Create output tensor for sampled token ids.
    if include_gpu_probs_tensor:
        sampled_token_ids_tensor = torch.empty(logprobs.shape[0],
                                               1,
                                               dtype=torch.long,
                                               device=logprobs.device)
    else:
        sampled_token_ids_tensor = None

476
477
    # Counterintiutively, having two loops here is actually faster.
    # The first loop can run without waiting on GPU<->CPU sync.
478
    for sampling_type in SamplingType:
479
        sample_indices = categorized_sample_indices[sampling_type][:, 0]
480
        num_tokens = len(sample_indices)
481
482
        if num_tokens == 0:
            continue
483

484
485
486
487
        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()
488
        if sampling_type == SamplingType.GREEDY:
489
            greedy_samples = torch.argmax(logprobs[long_sample_indices],
490
                                          dim=-1)
491
492
493
494
495
496
497
498
499
500
501
502
503
504

            if include_gpu_probs_tensor:
                # 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
505
        elif sampling_type in (SamplingType.RANDOM, SamplingType.RANDOM_SEED):
506
            max_best_of_in_batch = 1
507
508
509
            for seq_group in seq_groups:
                if seq_group.is_prompt:
                    sampling_params = seq_group.sampling_params
510
511
                    max_best_of_in_batch = max(max_best_of_in_batch,
                                               sampling_params.best_of)
Nick Hill's avatar
Nick Hill committed
512
513
514
            seeded_args = {} if sampling_type == SamplingType.RANDOM else {
                "seq_groups": seq_groups,
            }
515

Nick Hill's avatar
Nick Hill committed
516
            multinomial_samples[sampling_type] = _multinomial(
517
                probs[long_sample_indices], max_best_of_in_batch,
518
                **seeded_args)
519
520
521
522
523
524

            if include_gpu_probs_tensor:
                # Store sampled tokens in output tensor.
                sampled_token_ids_tensor[
                    long_sample_indices] = multinomial_samples[sampling_type]

525
526
527
528
529
530
        elif sampling_type == SamplingType.BEAM:
            beam_search_logprobs = logprobs[sample_indices]
        else:
            raise ValueError(f"Unsupported sampling type: {sampling_type}")

    # GPU<->CPU sync happens in the loop below.
531
    # This also converts the sample output to Python objects.
532
533
534
    for sampling_type in SamplingType:
        if sampling_type not in sample_metadata:
            continue
535
        (seq_group_id, seq_groups) = sample_metadata[sampling_type]
536
537
        if sampling_type == SamplingType.GREEDY:
            sample_results = _greedy_sample(seq_groups, greedy_samples)
Nick Hill's avatar
Nick Hill committed
538
        elif sampling_type in (SamplingType.RANDOM, SamplingType.RANDOM_SEED):
539
            sample_results = _random_sample(seq_groups,
Nick Hill's avatar
Nick Hill committed
540
                                            multinomial_samples[sampling_type])
541
        elif sampling_type == SamplingType.BEAM:
542
            sample_results = _beam_search_sample(seq_groups,
543
                                                 beam_search_logprobs)
544
        sample_results_dict.update(zip(seq_group_id, sample_results))
545

546
    sample_results = [
547
        sample_results_dict.get(i, ([], []))
548
        for i in range(len(sampling_metadata.seq_groups))
549
    ]
550
    return sample_results, sampled_token_ids_tensor
551
552


553
554
555
556
557
558
559
560
561
def _sample_with_triton_kernel(
    probs: torch.Tensor,
    logprobs: torch.Tensor,
    sampling_metadata: SamplingMetadata,
    sampling_tensors: SamplingTensors,
) -> List[Tuple[List[int], List[int]]]:
    categorized_seq_group_ids = {t: [] for t in SamplingType}
    categorized_sample_indices = sampling_metadata.categorized_sample_indices
    for i, seq_group in enumerate(sampling_metadata.seq_groups):
562
        sampling_params = seq_group.sampling_params
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
        sampling_type = sampling_params.sampling_type
        categorized_seq_group_ids[sampling_type].append(i)

    sample_results_dict: Dict[int, Tuple[List[int], List[int]]] = {}
    sample_metadata = {}
    max_best_of_in_batch = 1

    # Counterintiutively, having two loops here is actually faster.
    # The first loop can run without waiting on GPU<->CPU sync.
    for sampling_type in SamplingType:
        sample_indices = categorized_sample_indices[sampling_type][:, 0]
        sampled_token_indices = categorized_sample_indices[sampling_type][:, 1]
        num_tokens = len(sample_indices)
        if num_tokens == 0:
            continue
578
579
580
581
        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,
                                          sample_indices,
582
583
584
                                          sampled_token_indices)
        if sampling_type in (SamplingType.GREEDY, SamplingType.RANDOM,
                             SamplingType.RANDOM_SEED):
585
586
587
            for seq_group in seq_groups:
                if seq_group.is_prompt:
                    sampling_params = seq_group.sampling_params
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
                    max_best_of_in_batch = max(max_best_of_in_batch,
                                               sampling_params.best_of)
        elif sampling_type == SamplingType.BEAM:
            beam_search_logprobs = logprobs[sample_indices]
        else:
            raise ValueError(f"Unsupported sampling type: {sampling_type}")

    sampled_tokens, _, _ = sample_triton(
        probs=probs,
        seeds=sampling_tensors.sampling_seeds,
        max_best_of=max_best_of_in_batch,
        sample_indices=sampling_tensors.sample_indices,
        logprobs=logprobs,
        # don't save logprobs because we have logic for that below
        # TODO: use this instead of the CPU-based logic below
        save_logprobs=False,
    )

    # GPU<->CPU sync happens in the loop below.

    for sampling_type in SamplingType:
        if sampling_type not in sample_metadata:
            continue
611
        (seq_group_id, seq_groups, sample_indices,
612
613
614
615
616
617
         sampled_token_indices) = sample_metadata[sampling_type]
        if sampling_type == SamplingType.GREEDY:
            sample_results = _greedy_sample(
                seq_groups, sampled_tokens[sampled_token_indices][:, 0])
        elif sampling_type in (SamplingType.RANDOM, SamplingType.RANDOM_SEED):
            sample_results = _random_sample(
618
                seq_groups, sampled_tokens[sampled_token_indices])
619
        elif sampling_type == SamplingType.BEAM:
620
            sample_results = _beam_search_sample(seq_groups,
621
                                                 beam_search_logprobs)
622
        sample_results_dict.update(zip(seq_group_id, sample_results))
623
624

    sample_results = [
625
        sample_results_dict.get(i, ([], []))
626
627
628
629
630
631
        for i in range(len(sampling_metadata.seq_groups))
    ]
    return sample_results


def _sample(
632
633
634
635
    probs: torch.Tensor, logprobs: torch.Tensor,
    sampling_metadata: SamplingMetadata, sampling_tensors: SamplingTensors,
    include_gpu_probs_tensor: bool, modify_greedy_probs: bool
) -> Tuple[List[Tuple[List[int], List[int]]], Optional[torch.Tensor]]:
636
637
638
639
640
641
642
643
644
645
    """
    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 ([], [])
646
        sampled_token_ids_tensor: A tensor of sampled token ids.
647
    """
648
649
650
651
652
653
654
    return _sample_with_torch(
        probs,
        logprobs,
        sampling_metadata,
        include_gpu_probs_tensor=include_gpu_probs_tensor,
        modify_greedy_probs=modify_greedy_probs,
    )
655
656
657
658
659
660

    # TODO: Enable once Triton kernel & associated code is faster.
    # return _sample_with_triton_kernel(probs, logprobs, sampling_metadata,
    #                                   sampling_tensors)


661
def _get_ranks(x: torch.Tensor, indices: torch.Tensor) -> torch.Tensor:
662
663
664
665
666
667
    """
    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.
668
        indices (torch.Tensor): List of chosen token indices.
669
670
671
672
673
674

    Returns:
        torch.Tensor: 1D tensor of shape (N,) where N is the no. of tokens.
                    Each element in the returned tensor represents the rank 
                    of the chosen token in the input logprob tensor.
    """
675
676
677
    vals = x[torch.arange(0, len(x), device=x.device, dtype=indices.dtype),
             indices]
    return (x > vals[:, None]).long().sum(1).add_(1)
678
679


680
681
def _get_logprobs(
    logprobs: torch.Tensor,
682
    sampling_metadata: SamplingMetadata,
683
    sample_results: List[Tuple[List[int], List[int]]],
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
) -> Tuple[List[Optional[PromptLogprobs]], List[SampleLogprobs]]:
    """Return sample lobprobs and prompt logprobs.

    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.
    query_indices: List[int] = []
    # The next token ids to get the logprob value from.
    next_token_ids: List[int] = []
    # The largest requested number of logprobs. We find logprobs as many as the
    # largest num logprobs in this API.
718
    largest_num_logprobs = 1
719
720
721
722
723
724
725
726
727

    # 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
728
729
730
                and sampling_params.prompt_logprobs is not None):
            largest_num_logprobs = max(largest_num_logprobs,
                                       sampling_params.prompt_logprobs)
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
            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:
        empty_sampled_logprob = []
        empty_prompt_logprob = None
        return [empty_prompt_logprob], [empty_sampled_logprob]

    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,
766
    ]]
767
768
769
770
771
    ranks = _get_ranks(
        logprobs[query_indices_gpu],
        next_token_ids_gpu,
    )
    assert selected_logprobs.shape[0] == ranks.shape[0]
772

773
774
    # Logprobs of topk tokens for a batch of sequence groups.
    # (num_query_tokens_across_batch).
775
776
777
778
779
780
781
782
783
    if largest_num_logprobs > 0:
        top_logprobs, top_token_ids = torch.topk(logprobs,
                                                 largest_num_logprobs,
                                                 dim=-1)
        top_logprobs = top_logprobs.cpu()
        top_token_ids = top_token_ids.cpu()
    else:
        top_logprobs, top_token_ids = None, None

784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
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
834
835
836
    selected_logprobs = selected_logprobs.cpu()
    ranks = ranks.cpu()

    # Find prompt/sample logprobs.
    prompt_logprobs_per_seq_group: List[Optional[PromptLogprobs]] = []
    sample_logprobs_per_seq_group: List[SampleLogprobs] = []
    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
    if (is_prompt and sampling_params.prompt_logprobs is not None):
        prompt_logprobs = []
        num_logprobs = sampling_params.prompt_logprobs
        next_prompt_tokens = _get_next_prompt_tokens(seq_group)
        for token_id in next_prompt_tokens:
            # Calculate the prompt logprob of the real prompt tokens.
            # Use tuple here for performance (to use to_list()).
            # {token_id: (logprob, rank_from_vocab)}
            prompt_logprobs_dict: Dict[int, Tuple[float, int]] = {
                token_id: (selected_logprobs[selected_logprobs_idx].item(),
                           ranks[selected_logprobs_idx].item())
            }
837

838
839
840
841
842
            # Add top K prompt logprobs along with its rank.
            if num_logprobs > 0:
                prompt_logprobs_dict.update(
                    zip(
                        top_token_ids[top_logprob_idx, :num_logprobs].tolist(),
843
                        zip(
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
                            top_logprobs[
                                top_logprob_idx, :num_logprobs].tolist(),
                            # This is ranks. Since top_logprob is sorted,
                            # we can just use a range here.
                            range(1, num_logprobs + 1))))
            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
            selected_logprobs_idx += 1
    return prompt_logprobs, top_logprob_idx, selected_logprobs_idx


def _get_sampled_logprob_if_needed(
    seq_group: SequenceGroupToSample,
    sample_result: Tuple[List[int], List[int]],
    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
    num_logprobs = seq_group.sampling_params.logprobs
    if num_logprobs is None:
        num_logprobs = 0
    sampled_logprobs: SampleLogprobs = []
    next_token_ids, parent_seq_ids = sample_result

    if seq_group.do_sample:
        assert len(next_token_ids) > 0
        for (next_token_id, parent_id) in zip(next_token_ids, parent_seq_ids):
            # Calculate the sample logprob of the real sampled tokens.
            # Use tuple here for performance (to use to_list()).
            # token_id: (logprob, rank_from_vocab)
            sampled_logprobs_dict: Dict[int, Tuple[float, int]] = {
884
                next_token_id:
885
886
                (selected_logprobs[selected_logprobs_idx].item(),
                 ranks[selected_logprobs_idx].item())
887
            }
888
889
890
891
892
893
            # +1 to go to the next sampled token. Note that
            # selected_logprobs can contain duplicates unlike top_logprobs
            # when beam search is enabled.
            selected_logprobs_idx += 1

            # Second, add top K logprobs along with its rank.
894
            if num_logprobs >= 0:
895
                sampled_logprobs_dict.update(
896
                    zip(
897
                        top_token_ids[top_logprob_idx +
898
                                      parent_id, :num_logprobs].tolist(),
899
                        zip(
900
                            top_logprobs[top_logprob_idx +
901
                                         parent_id, :num_logprobs].tolist(),
902
903
                            # This is rank. Since top_logprob is sorted, we
                            # can just use a range here.
904
                            range(1, num_logprobs + 1))))
905
906
907
908
            sampled_logprobs.append({
                token_id: Logprob(*logprob_and_rank)
                for token_id, logprob_and_rank in
                sampled_logprobs_dict.items()
909
            })
910
911
912
913
        # There are len(seq_ids) number of sampled tokens for the current
        # sequence group in top_logprobs. Jump to the next seq_group.
        top_logprob_idx += len(seq_ids)
    return sampled_logprobs, top_logprob_idx, selected_logprobs_idx
914
915


916
917
918
919
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
955
956
957
958
959
960
961
962
963
964
965
966
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.
    
    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.
    """
    logprobs[sample_indices, :] = -float('inf')
    logprobs[sample_indices, greedy_samples] = 0.0
    probs[sample_indices, :] = 0
    probs[sample_indices, greedy_samples] = 1.0


967
968
def _build_sampler_output(
    sample_results: List[Tuple[List[int], List[int]]],
969
    sampling_metadata: SamplingMetadata,
970
971
    prompt_logprobs: List[Optional[PromptLogprobs]],
    sample_logprobs: List[SampleLogprobs],
972
    on_device_tensors: Optional[Tuple[torch.Tensor, torch.Tensor]],
973
) -> SamplerOutput:
974
975
976
977
978
979
980
981
982
    """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.
    """

983
984
    sampler_output = []
    for (seq_group, sample_result, group_prompt_logprobs,
985
         group_sample_logprobs) in zip(sampling_metadata.seq_groups,
986
987
                                       sample_results, prompt_logprobs,
                                       sample_logprobs):
988
        seq_ids = seq_group.seq_ids
989
990
991
992
993
994
        next_token_ids, parent_ids = sample_result
        seq_outputs = []
        for parent_id, next_token_id, logprobs in zip(parent_ids,
                                                      next_token_ids,
                                                      group_sample_logprobs):
            seq_outputs.append(
Zhuohan Li's avatar
Zhuohan Li committed
995
                SequenceOutput(seq_ids[parent_id], next_token_id, logprobs))
996
        sampler_output.append(
Zhuohan Li's avatar
Zhuohan Li committed
997
            SequenceGroupOutput(seq_outputs, group_prompt_logprobs))
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009

    # If not specified, store None values in SamplerOutput.
    if on_device_tensors is not None:
        sampled_token_probs, sampled_token_ids = on_device_tensors
    else:
        sampled_token_probs, sampled_token_ids = (None, None)

    return SamplerOutput(
        outputs=sampler_output,
        sampled_token_probs=sampled_token_probs,
        sampled_token_ids=sampled_token_ids,
    )
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042


def _get_next_prompt_tokens(seq_group: SequenceGroupToSample) -> List[str]:
    """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
    subquery_len = seq_group.subquery_len
    assert subquery_len is not None
    # 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
    next_token_index_end = min(computed_len + subquery_len + 1,
                               len(prompt_tokens))
    next_prompt_tokens = prompt_tokens[
        next_token_index_start:next_token_index_end]
    return next_prompt_tokens