flash_attn.py 39.8 KB
Newer Older
1
"""Attention layer with FlashAttention."""
2
from collections import defaultdict
3
from dataclasses import dataclass
4
from itertools import accumulate
5
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type
6
7
8

import torch

9
from vllm import _custom_ops as ops
10
from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
11
                                              AttentionLayer,
12
13
14
                                              AttentionMetadata,
                                              AttentionMetadataBuilder,
                                              AttentionType)
15
16
17
18
19
from vllm.attention.backends.utils import (
    PAD_SLOT_ID, CommonAttentionState, compute_slot_mapping,
    compute_slot_mapping_start_idx, get_num_prefill_decode_query_kv_tokens,
    get_seq_len_block_table_args, is_all_cross_attn_metadata_set,
    is_all_encoder_attn_metadata_set, is_block_tables_empty)
20
from vllm.envs import VLLM_FLASH_ATTN_VERSION
21
from vllm.multimodal import MultiModalPlaceholderMap
22
from vllm.platforms import current_platform
23
from vllm.utils import async_tensor_h2d, make_tensor_with_pad
24
25

if TYPE_CHECKING:
26
27
    from vllm.worker.model_runner import (ModelInputForGPUBuilder,
                                          ModelInputForGPUWithSamplingMetadata)
28

29
from vllm.vllm_flash_attn import (flash_attn_varlen_func,
30
31
                                  flash_attn_with_kvcache,
                                  is_fa_version_supported)
32
33


34
35
class FlashAttentionBackend(AttentionBackend):

36
37
    accept_output_buffer: bool = True

38
39
40
41
    @staticmethod
    def get_supported_head_sizes() -> List[int]:
        return [32, 64, 96, 128, 160, 192, 224, 256]

42
43
    @staticmethod
    def get_name() -> str:
44
        return "FLASH_ATTN"
45

46
47
48
49
50
    @staticmethod
    def get_impl_cls() -> Type["FlashAttentionImpl"]:
        return FlashAttentionImpl

    @staticmethod
51
52
    def get_metadata_cls() -> Type["AttentionMetadata"]:
        return FlashAttentionMetadata
53

54
55
56
57
    @staticmethod
    def get_builder_cls() -> Type["FlashAttentionMetadataBuilder"]:
        return FlashAttentionMetadataBuilder

58
59
60
61
    @staticmethod
    def get_state_cls() -> Type["CommonAttentionState"]:
        return CommonAttentionState

62
63
64
65
66
67
68
    @staticmethod
    def get_kv_cache_shape(
        num_blocks: int,
        block_size: int,
        num_kv_heads: int,
        head_size: int,
    ) -> Tuple[int, ...]:
69
70
71
        if block_size % 16 != 0:
            raise ValueError("Block size must be a multiple of 16.")
        return (2, num_blocks, block_size, num_kv_heads, head_size)
72
73
74
75
76

    @staticmethod
    def swap_blocks(
        src_kv_cache: torch.Tensor,
        dst_kv_cache: torch.Tensor,
77
        src_to_dst: torch.Tensor,
78
    ) -> None:
79
80
        src_key_cache = src_kv_cache[0]
        dst_key_cache = dst_kv_cache[0]
81
        ops.swap_blocks(src_key_cache, dst_key_cache, src_to_dst)
82
83
        src_value_cache = src_kv_cache[1]
        dst_value_cache = dst_kv_cache[1]
84
        ops.swap_blocks(src_value_cache, dst_value_cache, src_to_dst)
85
86
87
88

    @staticmethod
    def copy_blocks(
        kv_caches: List[torch.Tensor],
89
        src_to_dists: torch.Tensor,
90
    ) -> None:
91
92
        key_caches = [kv_cache[0] for kv_cache in kv_caches]
        value_caches = [kv_cache[1] for kv_cache in kv_caches]
93

94
        ops.copy_blocks(key_caches, value_caches, src_to_dists)
95
96
97


@dataclass
98
class FlashAttentionMetadata(AttentionMetadata):
99
100
101
102
103
104
105
    """Metadata for FlashAttentionBackend.

    NOTE: Any python object stored here is not updated when it is
    cuda-graph replayed. If you have values that need to be changed
    dynamically, it should be stored in tensor. The tensor has to be
    updated from `CUDAGraphRunner.forward` API.
    """
106
107
108
109
110
    # (batch_size,). The sequence length per sequence. Sequence length means
    # the computed tokens + new tokens None if it is a decoding.
    seq_lens: Optional[List[int]]
    # seq_lens stored as a tensor.
    seq_lens_tensor: Optional[torch.Tensor]
111

112
    # NOTE(sang): Definition of context_len, query_len, and seq_len.
113
114
115
116
    # |---------- N-1 iteration --------|
    # |---------------- N iteration ---------------------|
    # |- tokenA -|......................|-- newTokens ---|
    # |---------- context_len ----------|
117
    # |-------------------- seq_len ---------------------|
118
    #                                   |-- query_len ---|
119

120
121
122
123
124
125
    # Maximum sequence length among prefill batch. 0 if there are decoding
    # requests only.
    max_prefill_seq_len: int
    # Maximum sequence length among decode batch. 0 if there are prefill
    # requests only.
    max_decode_seq_len: int
126
127
128
    # (batch_size,) A tensor of context lengths (tokens that are computed
    # so far).
    context_lens_tensor: Optional[torch.Tensor]
129

130
131
132
133
134
135
136
137
    # (batch_size, max_blocks_per_seq).
    # Block addresses per sequence. (Seq id -> list of physical block)
    # E.g., [0, 1, 2] means tokens are stored in 0th, 1st, and 2nd blocks
    # in the kv cache. Each block can contain up to block_size tokens.
    # 2nd dimensions are padded up to max_blocks_per_seq if it is cuda-graph
    # captured.
    block_tables: Optional[torch.Tensor]

138
139
140
    # Whether or not if cuda graph is enabled.
    # Cuda-graph is currently enabled for decoding only.
    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
141

142
143
    use_cuda_graph: bool

144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
    # Maximum query length in the batch.
    max_query_len: Optional[int] = None

    # Max number of query tokens among request in the batch.
    max_decode_query_len: Optional[int] = None

    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
    # the batch, used to index into subquery. E.g., if the subquery length
    # is [4, 6], it is [0, 4, 10].
    query_start_loc: Optional[torch.Tensor] = None
    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
    # the batch, used to index into sequence. E.g., if the sequence length is
    # [4, 6], it is [0, 4, 10].
    seq_start_loc: Optional[torch.Tensor] = None

159
160
161
    _cached_prefill_metadata: Optional["FlashAttentionMetadata"] = None
    _cached_decode_metadata: Optional["FlashAttentionMetadata"] = None

162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
    # Begin encoder attn & enc/dec cross-attn fields...

    # Encoder sequence lengths representation
    encoder_seq_lens: Optional[List[int]] = None
    encoder_seq_lens_tensor: Optional[torch.Tensor] = None
    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
    # the batch, used to index into sequence. E.g., if the sequence length is
    # [4, 6], it is [0, 4, 10].
    encoder_seq_start_loc: Optional[torch.Tensor] = None
    # Maximum sequence length among encoder sequences
    max_encoder_seq_len: Optional[int] = None
    # Number of tokens input to encoder
    num_encoder_tokens: Optional[int] = None

    # Cross-attention memory-mapping data structures: slot mapping
    # and block tables
    cross_slot_mapping: Optional[torch.Tensor] = None
    cross_block_tables: Optional[torch.Tensor] = None

    @property
    def is_all_encoder_attn_metadata_set(self):
        '''
        All attention metadata required for encoder attention is set.
        '''
        return is_all_encoder_attn_metadata_set(self)

    @property
    def is_all_cross_attn_metadata_set(self):
        '''
        All attention metadata required for enc/dec cross-attention is set.

        Superset of encoder attention required metadata.
        '''
        return is_all_cross_attn_metadata_set(self)

197
198
199
200
201
202
203
204
    @property
    def prefill_metadata(self) -> Optional["FlashAttentionMetadata"]:
        if self.num_prefills == 0:
            return None

        if self._cached_prefill_metadata is not None:
            return self._cached_prefill_metadata

205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
        assert ((self.seq_lens is not None)
                or (self.encoder_seq_lens is not None))
        assert ((self.seq_lens_tensor is not None)
                or (self.encoder_seq_lens_tensor is not None))

        # Compute some attn_metadata fields which default to None
        query_start_loc = (None if self.query_start_loc is None else
                           self.query_start_loc[:self.num_prefills + 1])
        slot_mapping = (None if self.slot_mapping is None else
                        self.slot_mapping[:self.num_prefill_tokens])
        seq_lens = (None if self.seq_lens is None else
                    self.seq_lens[:self.num_prefills])
        seq_lens_tensor = (None if self.seq_lens_tensor is None else
                           self.seq_lens_tensor[:self.num_prefills])
        seq_start_loc = (None if self.seq_start_loc is None else
                         self.seq_start_loc[:self.num_prefills + 1])
        context_lens_tensor = (None if self.context_lens_tensor is None else
                               self.context_lens_tensor[:self.num_prefills])
        block_tables = (None if self.block_tables is None else
                        self.block_tables[:self.num_prefills])
225
226
227
228
229

        self._cached_prefill_metadata = FlashAttentionMetadata(
            num_prefills=self.num_prefills,
            num_prefill_tokens=self.num_prefill_tokens,
            num_decode_tokens=0,
230
            slot_mapping=slot_mapping,
231
232
            multi_modal_placeholder_index_maps=self.
            multi_modal_placeholder_index_maps,
233
            enable_kv_scales_calculation=self.enable_kv_scales_calculation,
234
235
            seq_lens=seq_lens,
            seq_lens_tensor=seq_lens_tensor,
236
237
            max_query_len=self.max_query_len,
            max_prefill_seq_len=self.max_prefill_seq_len,
238
            max_decode_query_len=0,
239
            max_decode_seq_len=0,
240
241
242
243
            query_start_loc=query_start_loc,
            seq_start_loc=seq_start_loc,
            context_lens_tensor=context_lens_tensor,
            block_tables=block_tables,
244
            use_cuda_graph=False,
245
246
247
248
249
250
251
            # Begin encoder & cross attn fields below...
            encoder_seq_lens=self.encoder_seq_lens,
            encoder_seq_lens_tensor=self.encoder_seq_lens_tensor,
            encoder_seq_start_loc=self.encoder_seq_start_loc,
            max_encoder_seq_len=self.max_encoder_seq_len,
            cross_slot_mapping=self.cross_slot_mapping,
            cross_block_tables=self.cross_block_tables)
252
253
254
255
256
257
258
259
260
        return self._cached_prefill_metadata

    @property
    def decode_metadata(self) -> Optional["FlashAttentionMetadata"]:
        if self.num_decode_tokens == 0:
            return None

        if self._cached_decode_metadata is not None:
            return self._cached_decode_metadata
261
262
263
264
265
266
267
268
269
270
        assert ((self.seq_lens_tensor is not None)
                or (self.encoder_seq_lens_tensor is not None))

        # Compute some attn_metadata fields which default to None
        slot_mapping = (None if self.slot_mapping is None else
                        self.slot_mapping[self.num_prefill_tokens:])
        seq_lens_tensor = (None if self.seq_lens_tensor is None else
                           self.seq_lens_tensor[self.num_prefills:])
        block_tables = (None if self.block_tables is None else
                        self.block_tables[self.num_prefills:])
271
272
273
274
275

        self._cached_decode_metadata = FlashAttentionMetadata(
            num_prefills=0,
            num_prefill_tokens=0,
            num_decode_tokens=self.num_decode_tokens,
276
            slot_mapping=slot_mapping,
277
            multi_modal_placeholder_index_maps=None,
278
            enable_kv_scales_calculation=True,
279
            seq_lens=None,
280
            seq_lens_tensor=seq_lens_tensor,
281
            max_decode_query_len=self.max_decode_query_len,
282
            max_query_len=self.max_query_len,
283
284
            max_prefill_seq_len=0,
            max_decode_seq_len=self.max_decode_seq_len,
285
286
287
288
289
            # Batch may be composed of prefill|decodes, adjust query start
            # indices to refer to the start of decodes. E.g.
            # in tokens:[3 prefills|6 decodes], query_start_loc=[3,9] => [0,6].
            query_start_loc=(self.query_start_loc[self.num_prefills:] -
                             self.query_start_loc[self.num_prefills])
290
291
292
            if self.query_start_loc is not None else None,
            seq_start_loc=self.seq_start_loc[self.num_prefills:]
            if self.seq_start_loc is not None else None,
293
            context_lens_tensor=None,
294
            block_tables=block_tables,
295
            use_cuda_graph=self.use_cuda_graph,
296
297
298
299
300
301
302
            # Begin encoder & cross attn fields below...
            encoder_seq_lens=self.encoder_seq_lens,
            encoder_seq_lens_tensor=self.encoder_seq_lens_tensor,
            encoder_seq_start_loc=self.encoder_seq_start_loc,
            max_encoder_seq_len=self.max_encoder_seq_len,
            cross_slot_mapping=self.cross_slot_mapping,
            cross_block_tables=self.cross_block_tables)
303
304
        return self._cached_decode_metadata

305
306
    def advance_step(self,
                     model_input: "ModelInputForGPUWithSamplingMetadata",
307
                     sampled_token_ids: Optional[torch.Tensor],
308
309
310
311
                     block_size: int,
                     num_seqs: int,
                     num_queries: int,
                     turn_prefills_into_decodes: bool = False):
312
313
314
315
316
317
318
319
320
321
        """
        Update metadata in-place to advance one decode step.
        """
        # When using cudagraph, the num_seqs is padded to the next captured
        # batch sized, but num_queries tracks the actual number of requests in
        # the batch. For --enforce-eager mode, num_seqs == num_queries
        if num_seqs != num_queries:
            assert num_seqs > num_queries
            assert self.use_cuda_graph

322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
        if turn_prefills_into_decodes:
            # When Mutli-Step is enabled with Chunked-Prefill, prefills and
            # decodes are scheduled together. In the first step, all the
            # prefills turn into decodes. This update reflects that
            # conversion.
            assert self.num_decode_tokens + self.num_prefills == num_seqs
            self.num_decode_tokens += self.num_prefills
            self.num_prefills = 0
            self.num_prefill_tokens = 0
            self.max_prefill_seq_len = 0
            self.max_query_len = 1

            self.slot_mapping = self.slot_mapping[:num_seqs]
        else:
            assert self.seq_lens is not None
            assert self.max_decode_seq_len == max(self.seq_lens)

339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
        assert self.num_prefills == 0
        assert self.num_prefill_tokens == 0
        assert self.num_decode_tokens == num_seqs
        assert self.slot_mapping.shape == (num_seqs, )

        assert self.seq_lens is not None
        assert len(self.seq_lens) == num_seqs
        assert self.seq_lens_tensor is not None
        assert self.seq_lens_tensor.shape == (num_seqs, )
        assert self.max_query_len == 1
        assert self.max_prefill_seq_len == 0

        assert self.query_start_loc is not None
        assert self.query_start_loc.shape == (num_queries + 1, )
        assert self.seq_start_loc is not None
        assert self.seq_start_loc.shape == (num_seqs + 1, )

        assert self.context_lens_tensor is not None
        assert self.context_lens_tensor.shape == (num_queries, )

        assert self.block_tables is not None
        assert self.block_tables.shape[0] == num_seqs

        # Update query lengths. Note that we update only queries and not seqs,
        # since tensors may be padded due to captured cuda graph batch size
        for i in range(num_queries):
            self.seq_lens[i] += 1
        self.max_decode_seq_len = max(self.seq_lens)

368
369
370
371
372
373
374
375
376
        ops.advance_step_flashattn(num_seqs=num_seqs,
                                   num_queries=num_queries,
                                   block_size=block_size,
                                   input_tokens=model_input.input_tokens,
                                   sampled_token_ids=sampled_token_ids,
                                   input_positions=model_input.input_positions,
                                   seq_lens=self.seq_lens_tensor,
                                   slot_mapping=self.slot_mapping,
                                   block_tables=self.block_tables)
377

378

379
380
381
382
class FlashAttentionMetadataBuilder(
        AttentionMetadataBuilder[FlashAttentionMetadata]):

    def __init__(self, input_builder: "ModelInputForGPUBuilder"):
383
384
385
386
387
388
        self.input_builder = input_builder
        self.runner = input_builder.runner
        self.sliding_window = input_builder.sliding_window
        self.block_size = input_builder.block_size

    def prepare(self):
389
390
391
392
393
        self.slot_mapping: List[int] = []
        self.prefill_seq_lens: List[int] = []
        self.context_lens: List[int] = []
        self.block_tables: List[List[int]] = []
        self.curr_seq_lens: List[int] = []
394
395
396
        self.multimodal_placeholder_maps: Dict[
            str,
            MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
397
398
399
        self.num_prefills = 0
        self.num_prefill_tokens = 0
        self.num_decode_tokens = 0
400
        self.has_prefix_cache_hit = False
401

402
403
    def _add_seq_group(
            self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
404
            chunked_prefill_enabled: bool, prefix_cache_hit: bool):
405
406
407
408
409
        """Add a sequence group to the metadata. Specifically update/append
        1. context length.
        2. block table.
        3. slot mapping.
        """
410
411
        is_prompt = inter_data.is_prompt
        block_tables = inter_data.block_tables
412
413
414

        for (seq_id, token_len, seq_len, curr_seq_len, query_len, context_len,
             curr_sliding_window_block) in zip(
415
416
417
418
                 inter_data.seq_ids, [len(t) for t in inter_data.input_tokens],
                 inter_data.orig_seq_lens, inter_data.seq_lens,
                 inter_data.query_lens, inter_data.context_lens,
                 inter_data.curr_sliding_window_blocks):
419
420
421
            self.context_lens.append(context_len)

            if is_prompt:
422
423
424
425
426
427
                mm_maps = inter_data.multi_modal_placeholder_maps
                if mm_maps:
                    for modality, placeholders in mm_maps.items():
                        self.multimodal_placeholder_maps[modality].extend(
                            placeholders)

428
429
430
431
432
433
434
435
436
437
438
439
                self.num_prefills += 1
                self.num_prefill_tokens += token_len
                self.prefill_seq_lens.append(seq_len)
            else:
                self.num_decode_tokens += query_len
                self.curr_seq_lens.append(curr_seq_len)

            # Compute block table.
            # TODO(sang): Combine chunked prefill and prefix caching by
            # only allowing multiple of block_size chunk size.
            # NOTE: This only works for oooooooxxx style attention.
            block_table = []
440
            if prefix_cache_hit:
441
442
443
444
445
                # NOTE(woosuk): For flash-attn, the block table should
                # include the entries for the incoming prefill tokens.
                block_table = block_tables[seq_id]
            elif ((chunked_prefill_enabled or not is_prompt)
                  and block_tables is not None):
446
447
448
449
450
                if curr_sliding_window_block == 0:
                    block_table = block_tables[seq_id]
                else:
                    block_table = block_tables[seq_id][
                        -curr_sliding_window_block:]
451
452
453
454
            self.block_tables.append(block_table)

            # Compute slot mapping.
            is_profile_run = is_block_tables_empty(block_tables)
455
456
457
            start_idx = compute_slot_mapping_start_idx(is_prompt, query_len,
                                                       context_len,
                                                       self.sliding_window)
458
459
            compute_slot_mapping(is_profile_run, self.slot_mapping, seq_id,
                                 seq_len, context_len, start_idx,
460
                                 self.block_size, inter_data.block_tables)
461

462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
    def _get_graph_runner_block_tables(
            self, num_seqs: int,
            block_tables: List[List[int]]) -> torch.Tensor:
        # The shape of graph_block_tables is
        # [max batch size, max context len // block size].
        max_batch_size, max_blocks = self.runner.graph_block_tables.shape
        assert max_batch_size >= num_seqs

        graph_block_tables = self.runner.graph_block_tables[:num_seqs]
        for i, block_table in enumerate(block_tables):
            if block_table:
                num_blocks = len(block_table)
                if num_blocks <= max_blocks:
                    graph_block_tables[i, :num_blocks] = block_table
                else:
                    # It may be possible to have more blocks allocated due
                    # to lookahead slots of multi-step, however, they are
                    # not used anyway, so can be safely ignored.
                    graph_block_tables[
                        i, :max_blocks] = block_table[:max_blocks]

        return torch.from_numpy(graph_block_tables).to(
            device=self.runner.device, non_blocking=True)

486
    def build(self, seq_lens: List[int], query_lens: List[int],
487
              cuda_graph_pad_size: int, batch_size: int):
488
489
490
491
492
493
494
495
496
        """Build attention metadata with on-device tensors.

        Args:
            seq_lens: The maybe padded sequence lengths of the input sequences.
            query_lens: The query lengths of the input sequences.
            cuda_graph_pad_size: The padding size for cuda graph.
                                 -1 if cuda graph is not used.
            batch_size: The maybe padded batch size.
        """
497
498
499
500
        prefix_cache_hit = any([
            inter_data.prefix_cache_hit
            for inter_data in self.input_builder.inter_data_list
        ])
501
502
        for inter_data in self.input_builder.inter_data_list:
            self._add_seq_group(inter_data,
503
504
                                self.input_builder.chunked_prefill_enabled,
                                prefix_cache_hit)
505
506

        device = self.runner.device
507
508
509
        use_captured_graph = cuda_graph_pad_size != -1

        max_query_len = max(query_lens)
510
511
        decode_query_lens = query_lens[self.num_prefills:]
        if len(decode_query_lens) > 0:
512
            max_decode_query_len = max(decode_query_lens)
513
        else:
514
            max_decode_query_len = 1
515
516
517
        max_prefill_seq_len = max(self.prefill_seq_lens, default=0)
        max_decode_seq_len = max(self.curr_seq_lens, default=0)
        num_decode_tokens = self.num_decode_tokens
518
519
        query_start_loc = list(accumulate(query_lens, initial=0))
        seq_start_loc = list(accumulate(seq_lens, initial=0))
520

521
        num_seqs = len(seq_lens)
522
523
524
        if use_captured_graph:
            self.slot_mapping.extend([PAD_SLOT_ID] * cuda_graph_pad_size)
            self.block_tables.extend([] * cuda_graph_pad_size)
525
526
527
            num_decode_tokens = batch_size - self.num_prefill_tokens
            block_tables = self._get_graph_runner_block_tables(
                num_seqs, self.block_tables)
528
529
530
531
532
533
534
535
536
        else:
            block_tables = make_tensor_with_pad(
                self.block_tables,
                pad=0,
                dtype=torch.int,
                device=device,
            )
        assert max_query_len > 0, ("query_lens: {}".format(query_lens))

537
538
539
540
541
542
543
        assert device is not None
        context_lens_tensor = async_tensor_h2d(self.context_lens, torch.int,
                                               device, self.runner.pin_memory)
        seq_lens_tensor = async_tensor_h2d(seq_lens, torch.int, device,
                                           self.runner.pin_memory)
        slot_mapping_tensor = async_tensor_h2d(self.slot_mapping, torch.long,
                                               device, self.runner.pin_memory)
544
545
546
547
548
        query_start_loc_tensor = async_tensor_h2d(query_start_loc, torch.int32,
                                                  device,
                                                  self.runner.pin_memory)
        seq_start_loc_tensor = async_tensor_h2d(seq_start_loc, torch.int32,
                                                device, self.runner.pin_memory)
549
550
551
552
553
        placeholder_index_maps = {
            modality: placeholder_map.index_map()
            for modality, placeholder_map in
            self.multimodal_placeholder_maps.items()
        }
554
555
556
557
558
559
560

        return FlashAttentionMetadata(
            num_prefills=self.num_prefills,
            slot_mapping=slot_mapping_tensor,
            num_prefill_tokens=self.num_prefill_tokens,
            num_decode_tokens=num_decode_tokens,
            seq_lens=seq_lens,
561
            multi_modal_placeholder_index_maps=placeholder_index_maps,
562
            enable_kv_scales_calculation=True,
563
564
            seq_lens_tensor=seq_lens_tensor,
            max_query_len=max_query_len,
565
            max_decode_query_len=max_decode_query_len,
566
567
            max_prefill_seq_len=max_prefill_seq_len,
            max_decode_seq_len=max_decode_seq_len,
568
569
            query_start_loc=query_start_loc_tensor,
            seq_start_loc=seq_start_loc_tensor,
570
571
572
573
574
575
            context_lens_tensor=context_lens_tensor,
            block_tables=block_tables,
            use_cuda_graph=use_captured_graph,
        )


576
577
578
class FlashAttentionImpl(AttentionImpl):
    """
    If the input tensors contain prompt tokens, the layout is as follows:
579
580
    |<--------------- num_prefill_tokens ----------------->|	
    |<--prefill_0-->|<--prefill_1-->|...|<--prefill_N-1--->|
581
582

    Otherwise, the layout is as follows:	
583
584
    |<----------------- num_decode_tokens ------------------>|	
    |<--decode_0-->|..........|<--decode_M-1-->|<--padding-->|
585
586
587
588
589
590

    Generation tokens can contain padding when cuda-graph is used.
    Currently, prompt tokens don't contain any padding.

    The prompts might have different lengths, while the generation tokens
    always have length 1.
591
592
593
594
595
596
597
598
599

    If chunked prefill is enabled, prefill tokens and decode tokens can be
    batched together in a flattened 1D query.

    |<----- num_prefill_tokens ---->|<------- num_decode_tokens --------->|
    |<-prefill_0->|...|<-prefill_N-1->|<--decode_0-->|...|<--decode_M-1-->|

    Currently, cuda graph is disabled for chunked prefill, meaning there's no
    padding between prefill and decode tokens.
600
601
602
603
604
605
606
    """

    def __init__(
        self,
        num_heads: int,
        head_size: int,
        scale: float,
607
608
609
610
        num_kv_heads: int,
        alibi_slopes: Optional[List[float]],
        sliding_window: Optional[int],
        kv_cache_dtype: str,
611
        blocksparse_params: Optional[Dict[str, Any]] = None,
612
        logits_soft_cap: Optional[float] = None,
613
        attn_type: str = AttentionType.DECODER,
614
    ) -> None:
615
616
617
        if blocksparse_params is not None:
            raise ValueError(
                "FlashAttention does not support block-sparse attention.")
618
619
620
        self.num_heads = num_heads
        self.head_size = head_size
        self.scale = float(scale)
621
        self.num_kv_heads = num_kv_heads
622
623
624
        if alibi_slopes is not None:
            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
        self.alibi_slopes = alibi_slopes
625
626
        self.sliding_window = ((sliding_window - 1,
                                0) if sliding_window is not None else (-1, -1))
627
        self.kv_cache_dtype = kv_cache_dtype
628
629
630
631
        if logits_soft_cap is None:
            # In flash-attn, setting logits_soft_cap as 0 means no soft cap.
            logits_soft_cap = 0
        self.logits_soft_cap = logits_soft_cap
632
633
634
635

        assert self.num_heads % self.num_kv_heads == 0
        self.num_queries_per_kv = self.num_heads // self.num_kv_heads

636
637
        support_head_sizes = FlashAttentionBackend.get_supported_head_sizes()
        if head_size not in support_head_sizes:
638
            raise ValueError(
639
                f"Head size {head_size} is not supported by FlashAttention. "
640
                f"Supported head sizes are: {support_head_sizes}.")
641
        self.attn_type = attn_type
642

643
644
645
646
647
648
649
650
651
652
653
654
655
656
        # if hopper default to FA3, otherwise stick to FA2 for now
        # TODO(lucas): profile FA3 on ampere to see if it makes sense to
        #  use FA3 as default for both
        if current_platform.get_device_capability()[0] >= 9:
            self.fa_version = 3 if is_fa_version_supported(3) else 2
        else:
            self.fa_version = 2

        if VLLM_FLASH_ATTN_VERSION is not None:
            assert VLLM_FLASH_ATTN_VERSION in [2, 3]
            self.fa_version = VLLM_FLASH_ATTN_VERSION

        assert is_fa_version_supported(self.fa_version)

657
658
    def forward(
        self,
659
        layer: AttentionLayer,
660
661
662
663
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        kv_cache: torch.Tensor,
664
        attn_metadata: FlashAttentionMetadata,
665
        output: Optional[torch.Tensor] = None,
666
    ) -> torch.Tensor:
667
        """Forward pass with FlashAttention.
668
669

        Args:
670
671
672
673
            query: shape = [num_tokens, num_heads, head_size]
            key: shape = [num_tokens, num_kv_heads, head_size]
            value: shape = [num_tokens, num_kv_heads, head_size]
            output: shape = [num_tokens, num_heads, head_size]
674
            kv_cache = [2, num_blocks, block_size, num_kv_heads, head_size]
675
676
                NOTE: kv_cache will be an empty tensor with shape [0]
                for profiling run.
677
            attn_metadata: Metadata for attention.
678
        NOTE: It in-place updates the output tensor.
679
        """
680
        # NOTE(woosuk): FlashAttention does not support FP8 KV cache.
681
        assert layer._k_scale_float == 1.0 and layer._v_scale_float == 1.0, (
682
            "key/v_scale is not supported in FlashAttention.")
683

684
685
        assert output is not None, "Output tensor must be provided."

686
        attn_type = self.attn_type
687
688
689
690
691
692
693
694
695
696
        if (attn_type == AttentionType.ENCODER
                and (not attn_metadata.is_all_encoder_attn_metadata_set)):
            raise AttributeError("Encoder attention requires setting "
                                 "encoder metadata attributes.")
        elif (attn_type == AttentionType.ENCODER_DECODER
              and (not attn_metadata.is_all_cross_attn_metadata_set)):
            raise AttributeError("Encoder/decoder cross-attention "
                                 "requires setting cross-attention "
                                 "metadata attributes.")

697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
        kv_cache_dtype: str = self.kv_cache_dtype
        softmax_scale: float = self.scale
        window_size = self.sliding_window
        alibi_slopes: Optional[torch.Tensor] = self.alibi_slopes
        logits_soft_cap: Optional[float] = self.logits_soft_cap

        if kv_cache.numel() > 0:
            key_cache = kv_cache[0]
            value_cache = kv_cache[1]
            # We skip updating the KV cache under two conditions:
            #  a. When the Attention Type is ENCODER. In this phase, we compute
            #     only the encoder attention without updating the cache.
            #  b. When both Key and Value are None. This occurs during
            #     cross-attention computation in the decoding phase, where the
            #     KV cache is already populated with the cross-attention
            #     tensor. Thus, we skip cache updates during this time.
            if (attn_type != AttentionType.ENCODER) and (key is not None) and (
                    value is not None):
                if attn_type == AttentionType.ENCODER_DECODER:
                    # Update cross-attention KV cache (prefill-only)
                    updated_slot_mapping = attn_metadata.cross_slot_mapping
                else:
                    # Update self-attention KV cache (prefill/decode)
                    updated_slot_mapping = attn_metadata.slot_mapping

                # Reshape the input keys and values and store them in the cache.
                # If kv_cache is not provided, the new key and value tensors are
                # not cached. This happens during the initial memory
                # profiling run.
                torch.ops._C_cache_ops.reshape_and_cache_flash(
                    key,
                    value,
                    kv_cache[0],
                    kv_cache[1],
                    updated_slot_mapping.flatten(),  # type: ignore[union-attr]
                    kv_cache_dtype,
733
734
                    layer._k_scale,
                    layer._v_scale,
735
736
737
738
739
740
                )

        (num_prefill_query_tokens, num_prefill_kv_tokens,
        num_decode_query_tokens) = \
            get_num_prefill_decode_query_kv_tokens(attn_metadata, attn_type)
        decode_query = query[num_prefill_query_tokens:]
741
        decode_output = output[num_prefill_query_tokens:]
742
743
        # QKV for prefill.
        query = query[:num_prefill_query_tokens]
744
        prefill_output = output[:num_prefill_query_tokens]
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
        assert query.shape[0] == num_prefill_query_tokens
        assert decode_query.shape[0] == num_decode_query_tokens

        if prefill_meta := attn_metadata.prefill_metadata:
            # Prompt run.
            if (kv_cache.numel() == 0 or prefill_meta.block_tables is None
                    or prefill_meta.block_tables.numel() == 0):
                # normal attention
                # When block_tables are not filled, it means q and k are the
                # prompt, and they have the same length.
                q_seq_start_loc, q_seq_len, k_seq_start_loc, k_seq_len = \
                    _get_query_key_seq_metadata(prefill_meta, True, attn_type)

                key = key[:num_prefill_kv_tokens]
                value = value[:num_prefill_kv_tokens]

761
                flash_attn_varlen_func(
762
763
764
765
766
767
768
769
770
771
772
773
                    q=query,
                    k=key,
                    v=value,
                    cu_seqlens_q=q_seq_start_loc,
                    cu_seqlens_k=k_seq_start_loc,
                    max_seqlen_q=q_seq_len,
                    max_seqlen_k=k_seq_len,
                    softmax_scale=softmax_scale,
                    causal=_get_causal_option(attn_type),
                    window_size=window_size,
                    alibi_slopes=alibi_slopes,
                    softcap=logits_soft_cap,
774
                    out=prefill_output,
775
                    fa_version=self.fa_version,
776
777
778
779
780
781
782
                )
            else:
                # prefix-enabled attention
                assert attn_type == AttentionType.DECODER, (
                    "Only decoder-only models support prefix caching")
                assert prefill_meta.seq_lens is not None
                max_seq_len = max(prefill_meta.seq_lens)
783
                flash_attn_varlen_func(  # noqa
784
785
786
787
788
                    q=query,
                    k=key_cache,
                    v=value_cache,
                    cu_seqlens_q=prefill_meta.query_start_loc,
                    max_seqlen_q=prefill_meta.max_query_len,
789
                    seqused_k=prefill_meta.seq_lens_tensor,
790
791
792
793
794
795
796
                    max_seqlen_k=max_seq_len,
                    softmax_scale=softmax_scale,
                    causal=True,
                    window_size=window_size,
                    alibi_slopes=alibi_slopes,
                    block_table=prefill_meta.block_tables,
                    softcap=logits_soft_cap,
797
                    out=prefill_output,
798
                    fa_version=self.fa_version,
799
800
801
802
803
804
805
806
807
808
809
810
811
                )

        if decode_meta := attn_metadata.decode_metadata:
            # Decoding run.
            # Use flash_attn_varlen_func kernel for speculative decoding
            # because different queries might have different lengths.

            assert decode_meta.max_decode_query_len is not None
            # use only for actual varlen decoding
            if decode_meta.max_decode_query_len > 1:
                assert attn_type == AttentionType.DECODER, (
                    "Only decoder-only models support max_decode_query_len > 1"
                )
812
                flash_attn_varlen_func(
813
814
815
816
817
                    q=decode_query,
                    k=key_cache,
                    v=value_cache,
                    cu_seqlens_q=decode_meta.query_start_loc,
                    max_seqlen_q=decode_meta.max_decode_query_len,
818
                    seqused_k=decode_meta.seq_lens_tensor,
819
820
821
822
823
824
825
                    max_seqlen_k=decode_meta.max_decode_seq_len,
                    softmax_scale=softmax_scale,
                    causal=True,
                    window_size=window_size,
                    alibi_slopes=alibi_slopes,
                    softcap=logits_soft_cap,
                    block_table=decode_meta.block_tables,
826
                    out=decode_output,
827
                    fa_version=self.fa_version,
828
829
830
831
832
833
834
835
                )
            else:
                # Use flash_attn_with_kvcache for normal decoding.
                (
                    seq_lens_arg,
                    _,
                    block_tables_arg,
                ) = get_seq_len_block_table_args(decode_meta, False, attn_type)
836
                flash_attn_with_kvcache(
837
838
839
840
841
842
843
844
845
846
                    q=decode_query.unsqueeze(1),
                    k_cache=key_cache,
                    v_cache=value_cache,
                    block_table=block_tables_arg,
                    cache_seqlens=seq_lens_arg,
                    softmax_scale=softmax_scale,
                    causal=True,
                    window_size=window_size,
                    alibi_slopes=alibi_slopes,
                    softcap=logits_soft_cap,
847
                    out=decode_output.unsqueeze(1),
848
                    fa_version=self.fa_version,
849
                )
850
851
852
        return output


853
854
855
def _get_query_key_seq_metadata(
    attn_metadata,
    is_prompt: bool,
856
    attn_type: str,
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
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
913
914
915
916
917
) -> tuple:
    """
    Returns sequence metadata for key and query based on the specified 
    attention type and whether input is a prompt.

    This function computes the starting locations and maximum sequence lengths 
    for key and query sequences for different attention types.

    Args:
        attn_metadata: The attention metadata object
        is_prompt (bool): A flag indicating if the input is a prompt
        attn_type (AttentionType): The type of attention being used.

    Returns:
        tuple: A tuple containing four integers:
            - Starting location for the query sequence.
            - Maximum sequence length for the query sequence.
            - Starting location for the key sequence.
            - Maximum sequence length for the key sequence.

    Raises:
        AttributeError: If an invalid attention type is provided.
    """
    if attn_type == AttentionType.DECODER:
        # Decoder self-attention
        # Choose max_seq_len based on whether we are in prompt_run
        if is_prompt:
            max_seq_len = attn_metadata.max_prefill_seq_len
        else:
            max_seq_len = attn_metadata.max_decode_seq_len
        return (attn_metadata.seq_start_loc, max_seq_len,
                attn_metadata.seq_start_loc, max_seq_len)

    elif attn_type == AttentionType.ENCODER_DECODER:
        # This is cross attention between the where the key
        # is the precomputed encoder attention and query
        # is the input sequence.
        # Choose query max length based on whether it is prompt
        # or not.
        if is_prompt:
            max_seq_len = attn_metadata.max_prefill_seq_len
        else:
            max_seq_len = attn_metadata.max_decode_seq_len
        return (attn_metadata.seq_start_loc, max_seq_len,
                attn_metadata.encoder_seq_start_loc,
                attn_metadata.max_encoder_seq_len)
    elif attn_type == AttentionType.ENCODER:
        # For encoder attention both the query and the key are same i.e the
        # encoder sequence.
        return (attn_metadata.encoder_seq_start_loc,
                attn_metadata.max_encoder_seq_len,
                attn_metadata.encoder_seq_start_loc,
                attn_metadata.max_encoder_seq_len)
    elif attn_type == AttentionType.ENCODER_ONLY:
        assert is_prompt, "Should not have decode for encoder only model."
        return (attn_metadata.seq_start_loc, attn_metadata.max_prefill_seq_len,
                attn_metadata.seq_start_loc, attn_metadata.max_prefill_seq_len)
    else:
        raise AttributeError(f"Invalid attention type {str(attn_type)}")


918
def _get_causal_option(attn_type: str) -> bool:
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
    """
    Determine whether the given attention type is suitable for causal 
    attention mechanisms.

    Args:
        attn_type (AttentionType): The type of attention being evaluated

    Returns:
        bool: Returns `True` if the attention type is suitable for causal 
        attention (i.e., not encoder, encoder-only, or encoder-decoder), 
        otherwise returns `False`.
    """
    return not (attn_type == AttentionType.ENCODER
                or attn_type == AttentionType.ENCODER_ONLY
                or attn_type == AttentionType.ENCODER_DECODER)