model_runner.py 47.9 KB
Newer Older
1
import time
2
3
from enum import IntEnum
from typing import Dict, List, NamedTuple, Optional, Set, Tuple
4

5
import numpy as np
6
import torch
7
import torch.nn as nn
8

9
10
from vllm.attention import (AttentionMetadata, AttentionMetadataPerStage,
                            get_attn_backend)
11
12
13
from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
                         ModelConfig, ParallelConfig, SchedulerConfig,
                         VisionLanguageConfig)
14
15
16
from vllm.distributed import broadcast_tensor_dict
from vllm.distributed.communication_op import graph_capture_mode
from vllm.distributed.device_communicators import custom_all_reduce
17
from vllm.logger import init_logger
18
19
20
from vllm.lora.layers import LoRAMapping
from vllm.lora.request import LoRARequest
from vllm.lora.worker_manager import LRUCacheWorkerLoRAManager
21
from vllm.model_executor import SamplingMetadata
22
from vllm.model_executor.model_loader import get_model
23
from vllm.sampling_params import SamplingParams
24
25
from vllm.sequence import (MultiModalData, SamplerOutput, SequenceData,
                           SequenceGroupMetadata)
26
27
from vllm.utils import (CudaMemoryProfiler, get_kv_cache_torch_dtype, is_hip,
                        is_pin_memory_available, make_tensor_with_pad)
28
29
30
31

logger = init_logger(__name__)

_PAD_SLOT_ID = -1
32
LORA_WARMUP_RANK = 8
33
34
_BATCH_SIZE_ALIGNMENT = 8
# Capture graphs for token size 1, 2, 4, 8, 16, 24, 32, 40, ..., 256.
35
# NOTE: _get_graph_batch_size needs to be updated if this list is changed.
36
37
38
_BATCH_SIZES_TO_CAPTURE = [1, 2, 4] + [
    _BATCH_SIZE_ALIGNMENT * i for i in range(1, 33)
]
39
40


41
42
43
44
class PreparePromptMetadata(NamedTuple):
    input_tokens: List[int]
    input_positions: List[int]
    attn_metadata: Optional[AttentionMetadataPerStage]
45
46
    seq_lens: List[int]
    query_lens: List[int]
47
48
49
50
51
52
53
54
55
56
57
58
    lora_index_mapping: List[int]
    lora_prompt_mapping: List[int]
    lora_requests: Set[LoRARequest]
    multi_modal_input: Optional[torch.Tensor]
    slot_mapping: List[int]

    @classmethod
    def empty(cls):
        return PreparePromptMetadata(
            input_tokens=[],
            input_positions=[],
            attn_metadata=None,
59
60
            seq_lens=[],
            query_lens=[],
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
            lora_index_mapping=[],
            lora_prompt_mapping=[],
            lora_requests=set(),
            multi_modal_input=None,
            slot_mapping=[],
        )


class PrepareDecodeMetadata(NamedTuple):
    input_tokens: List[int]
    input_positions: List[int]
    attn_metadata: Optional[AttentionMetadata]
    lora_index_mapping: List[int]
    lora_prompt_mapping: List[int]
    lora_requests: Set[LoRARequest]
    slot_mapping: List[int]

    @classmethod
    def empty(cls):
        return PrepareDecodeMetadata(
            input_tokens=[],
            input_positions=[],
            attn_metadata=None,
            lora_index_mapping=[],
            lora_prompt_mapping=[],
            lora_requests=set(),
            slot_mapping=[],
        )


# How batches are constructed.
class BatchType(IntEnum):
    # Every batch is prefill.
    PREFILL = 0
    # Every batch is decode.
    DECODE = 1
    # Batch is a mixture of prefill and decode.
    MIXED = 2


101
102
103
104
105
106
107
class ModelRunner:

    def __init__(
        self,
        model_config: ModelConfig,
        parallel_config: ParallelConfig,
        scheduler_config: SchedulerConfig,
108
        device_config: DeviceConfig,
109
        cache_config: CacheConfig,
110
        load_config: LoadConfig,
111
        lora_config: Optional[LoRAConfig],
112
        kv_cache_dtype: Optional[str] = "auto",
113
        is_driver_worker: bool = False,
114
        vision_language_config: Optional[VisionLanguageConfig] = None,
115
116
117
118
    ):
        self.model_config = model_config
        self.parallel_config = parallel_config
        self.scheduler_config = scheduler_config
119
120
        self.device_config = device_config
        self.cache_config = cache_config
121
        self.lora_config = lora_config
122
        self.load_config = load_config
123
        self.is_driver_worker = is_driver_worker
124
        self.vision_language_config = vision_language_config
125

126
        self.device = self.device_config.device
127
        self.pin_memory = is_pin_memory_available()
128

129
130
131
132
        self.kv_cache_dtype = kv_cache_dtype
        self.sliding_window = model_config.get_sliding_window()
        self.block_size = cache_config.block_size
        self.max_seq_len_to_capture = self.model_config.max_seq_len_to_capture
133
        self.graph_runners: Dict[int, CUDAGraphRunner] = {}
134
135
136
        self.graph_memory_pool: Optional[Tuple[
            int, int]] = None  # Set during graph capture.
        # When using CUDA graph, the input block tables must be padded to
137
        # max_seq_len_to_capture. However, creating the block table in
138
139
140
141
        # Python can be expensive. To optimize this, we cache the block table
        # in numpy and only copy the actual input content at every iteration.
        # The shape of the cached block table will be
        # (max batch size to capture, max context len to capture / block size).
142
143
144
145
        self.graph_block_tables = np.zeros(
            (max(_BATCH_SIZES_TO_CAPTURE), self.get_max_block_per_batch()),
            dtype=np.int32)
        self.attn_backend = get_attn_backend(self.model_config.dtype)
146

147
148
        # Lazy initialization
        self.model: torch.nn.Module  # Set after load_model
149
150
        # Set if the backend is flashinfer.
        self.flashinfer_workspace_buffer: torch.Tensor
151
152
        # Set after load_model.
        self.lora_manager: Optional[LRUCacheWorkerLoRAManager] = None
153

154
    def load_model(self) -> None:
155
        with CudaMemoryProfiler() as m:
156
            self.model = get_model(
157
158
159
                model_config=self.model_config,
                device_config=self.device_config,
                load_config=self.load_config,
160
161
162
                lora_config=self.lora_config,
                vision_language_config=self.vision_language_config,
                parallel_config=self.parallel_config,
163
164
                scheduler_config=self.scheduler_config,
            )
165
166

        self.model_memory_usage = m.consumed_memory
167
168
        logger.info("Loading model weights took %.4f GB",
                    self.model_memory_usage / float(2**30))
169
170

        if self.lora_config:
171
172
173
            assert hasattr(self.model, "supported_lora_modules"
                           ) and self.model.supported_lora_modules, (
                               "Model does not support LoRA")
Terry's avatar
Terry committed
174
175
176
177
178
            assert hasattr(
                self.model,
                "embedding_modules"), "Model does not have embedding_modules"
            assert hasattr(self.model, "embedding_padding_modules"
                           ), "Model does not have embedding_padding_modules"
179
180
            self.lora_manager = LRUCacheWorkerLoRAManager(
                self.scheduler_config.max_num_seqs,
181
                self.scheduler_config.max_num_batched_tokens, self.vocab_size,
Terry's avatar
Terry committed
182
183
                self.lora_config, self.device, self.model.embedding_modules,
                self.model.embedding_padding_modules)
184
            self.model = self.lora_manager.create_lora_manager(self.model)
185

186
187
188
189
190
191
192
        if self.kv_cache_dtype == "fp8" and is_hip():
            # Currently scaled KV cache is only enabled on ROCm
            if self.model_config.quantization_param_path is not None:
                if callable(getattr(self.model, "load_kv_cache_scales", None)):
                    self.model.load_kv_cache_scales(
                        self.model_config.quantization_param_path)
                else:
193
194
195
196
                    raise RuntimeError(
                        "Using FP8 KV cache and scaling factors provided but "
                        "model %s does not support loading scaling factors.",
                        self.model.__class__)
197
            else:
198
199
200
201
                logger.warning(
                    "Using FP8 KV cache but no scaling factors "
                    "provided. Defaulting to scaling factors of 1.0. "
                    "This may lead to less accurate results!")
202
        elif self.model_config.quantization_param_path is not None:
203
204
205
            logger.warning("KV cache scaling factors provided, "
                           "but the KV cache data type is not FP8. "
                           "KV cache scaling factors will not be used.")
206

207
208
    def get_max_block_per_batch(self) -> int:
        block_size = self.block_size
209
        return (self.max_seq_len_to_capture + block_size - 1) // block_size
210

211
212
213
    def _prepare_prompt(
        self,
        seq_group_metadata_list: List[SequenceGroupMetadata],
214
    ) -> PreparePromptMetadata:
215
216
217
        input_tokens: List[int] = []
        input_positions: List[int] = []
        slot_mapping: List[int] = []
218
219
220
        lora_index_mapping: List[int] = []
        lora_prompt_mapping: List[int] = []
        lora_requests: Set[LoRARequest] = set()
221

222
        seq_lens: List[int] = []
223
        context_lens: List[int] = []
224
        query_lens: List[int] = []
225
        prefix_block_tables: List[List[int]] = []
226
        multi_modal_input_list: List[torch.Tensor] = []
227

228
229
230
        if len(seq_group_metadata_list) == 0:
            return PreparePromptMetadata.empty()

231
232
233
234
235
236
        for seq_group_metadata in seq_group_metadata_list:
            assert seq_group_metadata.is_prompt
            seq_ids = list(seq_group_metadata.seq_data.keys())
            assert len(seq_ids) == 1
            seq_id = seq_ids[0]

237
238
239
            computed_block_nums = seq_group_metadata.computed_block_nums
            if (self.scheduler_config is not None
                    and self.scheduler_config.chunked_prefill_enabled
240
241
                    and not (computed_block_nums is None
                             or computed_block_nums == [])):
242
243
244
245
246
                raise RuntimeError(
                    "chunked prefill cannot be used with prefix caching "
                    "now.")

            token_chunk_size = seq_group_metadata.token_chunk_size
247
            seq_data = seq_group_metadata.seq_data[seq_id]
248
            context_len = seq_data.get_num_computed_tokens()
249
250
            # We should use get_len here because in case of preemption
            # it contains output tokens.
251
252
253
            seq_len = min(seq_data.get_len(), context_len + token_chunk_size)
            prompt_tokens = seq_data.get_token_ids()[context_len:seq_len]
            seq_lens.append(seq_len)
254
255
256
257
258

            # NOTE: This only works for oooooooxxx style attention.
            if computed_block_nums is not None and len(
                    computed_block_nums) > 0 and self.sliding_window is None:
                # Prefix is not supported with sliding_window
259
260
                context_len = len(computed_block_nums) * self.block_size
                prompt_tokens = prompt_tokens[context_len:]
261
                prefix_block_tables.append(computed_block_nums)
262
263
264
265
266
267
268
269
            elif self.scheduler_config.chunked_prefill_enabled:
                if seq_group_metadata.block_tables is not None:
                    # Prefill has chunked before.
                    block_table = seq_group_metadata.block_tables[seq_id]
                    prefix_block_tables.append(block_table)
                else:
                    # The first prefill.
                    prefix_block_tables.append([])
270
271
            else:
                prefix_block_tables.append([])
272
273
                # Right now, prefill start is always 0. However, this
                # assumption can be changed once chunked prefill is introduced.
274
                assert context_len == 0
275

276
            # actual prompt lens
277
278
            context_lens.append(context_len)
            query_lens.append(seq_len - context_len)
279

280
            input_tokens.extend(prompt_tokens)
281
282
            # NOTE(woosuk): Here we assume that the first token in the prompt
            # is always the first token in the sequence.
283
            input_positions.extend(list(range(context_len, seq_len)))
284
285
286
287
288
            lora_id = seq_group_metadata.lora_int_id

            if lora_id > 0:
                lora_requests.add(seq_group_metadata.lora_request)

289
            lora_index_mapping += [lora_id] * (seq_len - context_len)
290
291
            lora_prompt_mapping.extend(
                [lora_id] *
292
                (seq_len - context_len
293
294
                 if seq_group_metadata.sampling_params.prompt_logprobs else 1))

295
296
297
298
            if seq_group_metadata.multi_modal_data:
                multi_modal_input_list.append(
                    seq_group_metadata.multi_modal_data.data)

299
300
301
            if seq_group_metadata.block_tables is None:
                # During memory profiling, the block tables are not initialized
                # yet. In this case, we just use a dummy slot mapping.
302
                slot_mapping.extend([_PAD_SLOT_ID] * seq_len)
303
304
305
306
                continue

            # Compute the slot mapping.
            block_table = seq_group_metadata.block_tables[seq_id]
307

308
            # Mask the [0, start_idx) tokens of the prompt with _PAD_SLOT_ID,
309
            # where start_idx is max(0, seq_len - sliding_window).
310
311
312
313
314
            # For example, if the prompt len is 10, sliding window is 8, and
            # block size is 4, the first two tokens are masked and the slot
            # mapping will be [-1, -1, 2, 3, 4, 5, 6, 7, 0, 1].
            start_idx = 0
            if self.sliding_window is not None:
315
                assert context_len == 0, (
316
317
                    "Prefix caching is currently not supported with "
                    "sliding window attention")
318
                start_idx = max(0, seq_len - self.sliding_window)
319

320
            for i in range(context_len, seq_len):
321
                if i < start_idx:
322
                    slot_mapping.append(_PAD_SLOT_ID)
323
324
325
326
327
                    continue

                block_number = block_table[i // self.block_size]
                block_offset = i % self.block_size
                slot = block_number * self.block_size + block_offset
328
329
                slot_mapping.append(slot)

330
331
332
        max_query_len = max(query_lens)
        max_seq_len = max(seq_lens)
        assert max_query_len > 0
333

334
335
        context_lens_tensor = torch.tensor(context_lens,
                                           dtype=torch.int,
336
                                           device=self.device)
337
338
339
340
341
342
343
344
345
346

        if multi_modal_input_list:
            assert self.vision_language_config, (
                "Multi-modal inputs are only supported by "
                "vision language models.")
            multi_modal_input = torch.cat(multi_modal_input_list,
                                          dim=0).to(self.device)
        else:
            multi_modal_input = None

347
348
        # Prepare prefix block tables
        max_prompt_block_table_len = max(len(t) for t in prefix_block_tables)
349
        block_tables = make_tensor_with_pad(
350
351
352
353
            prefix_block_tables,
            max_len=max_prompt_block_table_len,
            pad=0,
            dtype=torch.int,
354
            device=self.device,
355
        )
356
357
358

        # Query length can be shorter than key (i.e., prompt) when prefill
        # is chunked or prefix cached.
359
360
361
362
        query_lens_tensor = torch.tensor(query_lens,
                                         dtype=torch.long,
                                         device=self.device)
        subquery_start_loc = torch.zeros(query_lens_tensor.shape[0] + 1,
363
364
365
                                         dtype=torch.int32,
                                         device=self.device)

366
367
368
369
        seq_lens_tensor = torch.tensor(seq_lens,
                                       dtype=torch.int,
                                       device=self.device)
        seq_start_loc = torch.zeros(seq_lens_tensor.shape[0] + 1,
370
371
372
                                    dtype=torch.int32,
                                    device=self.device)

373
        torch.cumsum(query_lens_tensor,
374
375
376
377
                     dim=0,
                     dtype=subquery_start_loc.dtype,
                     out=subquery_start_loc[1:])

378
        torch.cumsum(seq_lens_tensor,
379
380
381
                     dim=0,
                     dtype=seq_start_loc.dtype,
                     out=seq_start_loc[1:])
382

383
        if self.attn_backend.get_name() == "flashinfer":
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
            attn_metadata = self.attn_backend.make_metadata(
                is_prompt=True,
                use_cuda_graph=False,
                seq_start_loc=seq_start_loc,
                max_seq_len=max_seq_len,
                block_tables=block_tables)
        else:
            attn_metadata = self.attn_backend.make_metadata(
                is_prompt=True,
                seq_lens=seq_lens,
                seq_lens_tensor=seq_lens_tensor,
                max_query_len=max_query_len,
                max_seq_len=max_seq_len,
                subquery_start_loc=subquery_start_loc,
                seq_start_loc=seq_start_loc,
                context_lens_tensor=context_lens_tensor,
                block_tables=block_tables,
                use_cuda_graph=False,
            )
403
404
405
406
407

        return PreparePromptMetadata(
            input_tokens=input_tokens,
            input_positions=input_positions,
            attn_metadata=attn_metadata,
408
409
            seq_lens=seq_lens,
            query_lens=query_lens,
410
411
412
413
414
415
            lora_index_mapping=lora_index_mapping,
            lora_prompt_mapping=lora_prompt_mapping,
            lora_requests=lora_requests,
            multi_modal_input=multi_modal_input,
            slot_mapping=slot_mapping,
        )
416
417
418
419

    def _prepare_decode(
        self,
        seq_group_metadata_list: List[SequenceGroupMetadata],
420
    ) -> PrepareDecodeMetadata:
421
422
423
        input_tokens: List[int] = []
        input_positions: List[int] = []
        slot_mapping: List[int] = []
424
        seq_lens: List[int] = []
425
        block_tables: List[List[int]] = []
426
427
428
        lora_index_mapping: List[int] = []
        lora_prompt_mapping: List[int] = []
        lora_requests: Set[LoRARequest] = set()
429

430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
        # The following fields are only for flashinfer
        # Please follow https://docs.flashinfer.ai/tutorials/kv_layout.html#page-layout
        # for the precise definition of the following fields.
        # An example:
        # request 1, page indices [0, 5, 8]
        # request 2, page indices [1, 6, 7]
        # request 3, page indices [3, 4]
        # paged_kv_indices is a concatenation of page indices of all requests:
        # [0, 5, 8, 1, 6, 7, 3, 4]
        # paged_kv_indptr is used to index into paged_kv_indices:
        # [0, 3, 6, 8]
        paged_kv_indices: List[int] = []
        # 0 at the beginning of paged_kv_indptr indicates the start of the
        # first request’s page indices in the paged_kv_indices list.
        paged_kv_indptr: List[int] = [0]
        # paged_kv_last_page_len is the length of the last page of each request
        paged_kv_last_page_len: List[int] = []

448
449
450
        if len(seq_group_metadata_list) == 0:
            return PrepareDecodeMetadata.empty()

451
452
        for seq_group_metadata in seq_group_metadata_list:
            assert not seq_group_metadata.is_prompt
453
            assert seq_group_metadata.token_chunk_size == 1
454
455

            seq_ids = list(seq_group_metadata.seq_data.keys())
456
457
458
459
460
            lora_id = seq_group_metadata.lora_int_id

            if lora_id > 0:
                lora_requests.add(seq_group_metadata.lora_request)

461
462
463
            for seq_id in seq_ids:
                seq_data = seq_group_metadata.seq_data[seq_id]
                generation_token = seq_data.get_last_token_id()
464
                input_tokens.append(generation_token)
465

466
467
                seq_len = seq_data.get_len()
                position = seq_len - 1
468
                input_positions.append(position)
469

470
                seq_len = seq_len if self.sliding_window is None else min(
471
                    seq_len, self.sliding_window)
472
                seq_lens.append(seq_len)
473

474
475
476
477
                block_table = seq_group_metadata.block_tables[seq_id]
                block_number = block_table[position // self.block_size]
                block_offset = position % self.block_size
                slot = block_number * self.block_size + block_offset
478
479
                slot_mapping.append(slot)
                lora_index_mapping.append(lora_id)
480
                lora_prompt_mapping.append(lora_id)
481
482
483
484
485
486
487

                if self.sliding_window is not None:
                    sliding_window_blocks = (self.sliding_window //
                                             self.block_size)
                    block_table = block_table[-sliding_window_blocks:]
                block_tables.append(block_table)

488
489
490
491
492
493
494
                paged_kv_indices.extend(block_table)
                paged_kv_indptr.append(paged_kv_indptr[-1] + len(block_table))
                last_page_len = seq_data.get_len() % self.block_size
                if last_page_len == 0:
                    last_page_len = self.block_size
                paged_kv_last_page_len.append(last_page_len)

495
496
497
        # vLLM uses cuda graph only for decoding requests.
        # See `capture_model` API for more details.
        # For decoding requests, batch_size == input_tokens.
498
        batch_size = len(input_tokens)
499
500
501
502
        max_seq_len = max(seq_lens)
        use_captured_graph = (not self.model_config.enforce_eager
                              and batch_size <= _BATCH_SIZES_TO_CAPTURE[-1]
                              and max_seq_len <= self.max_seq_len_to_capture)
503
504
505
506
        if use_captured_graph:
            graph_batch_size = _get_graph_batch_size(batch_size)
            assert graph_batch_size >= batch_size
            for _ in range(graph_batch_size - batch_size):
507
508
509
                input_tokens.append(0)
                input_positions.append(0)
                slot_mapping.append(_PAD_SLOT_ID)
510
                seq_lens.append(1)
511
                block_tables.append([])
512
                lora_index_mapping.append(0)
513
514
            batch_size = graph_batch_size

515
516
517
        seq_lens_tensor = torch.tensor(seq_lens,
                                       dtype=torch.int,
                                       device=self.device)
518
519

        if use_captured_graph:
520
521
            # When using cuda-graph all these tensors should be
            # padded.
522
523
524
            assert seq_lens_tensor.shape[0] == len(input_tokens)
            assert seq_lens_tensor.shape[0] == len(input_positions)
            assert seq_lens_tensor.shape[0] == len(slot_mapping)
525

526
527
528
529
530
531
            # The shape of graph_block_tables is
            # [max batch size, max context len // block size].
            input_block_tables = self.graph_block_tables[:batch_size]
            for i, block_table in enumerate(block_tables):
                if block_table:
                    input_block_tables[i, :len(block_table)] = block_table
532
            block_tables = torch.tensor(input_block_tables, device=self.device)
533
        else:
534
535
            max_block_table_len = max(
                len(block_table) for block_table in block_tables)
536
            block_tables = make_tensor_with_pad(
537
                block_tables,
538
                max_len=max_block_table_len,
539
540
                pad=0,
                dtype=torch.int,
541
                device=self.device,
542
            )
543

544
        if self.attn_backend.get_name() == "flashinfer":
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
            if not hasattr(self, "flashinfer_workspace_buffer"):
                # Allocate 16MB workspace buffer
                # Follow the example of flashinfer: https://docs.flashinfer.ai/api/python/decode.html
                self.flashinfer_workspace_buffer = torch.empty(
                    16 * 1024 * 1024, dtype=torch.uint8, device=self.device)
            paged_kv_indptr = torch.tensor(paged_kv_indptr,
                                           dtype=torch.int,
                                           device=self.device)
            paged_kv_indices = torch.tensor(paged_kv_indices,
                                            dtype=torch.int,
                                            device=self.device)
            paged_kv_last_page_len = torch.tensor(paged_kv_last_page_len,
                                                  dtype=torch.int,
                                                  device=self.device)
            kv_cache_dtype = get_kv_cache_torch_dtype(self.kv_cache_dtype,
                                                      self.model_config.dtype)

            attn_metadata = self.attn_backend.make_metadata(
                is_prompt=False,
                use_cuda_graph=False,
                workspace_buffer=self.flashinfer_workspace_buffer,
                paged_kv_indptr=paged_kv_indptr,
                paged_kv_indices=paged_kv_indices,
                paged_kv_last_page_len=paged_kv_last_page_len,
                num_qo_heads=self.model_config.get_num_attention_heads(
                    self.parallel_config),
                num_kv_heads=self.model_config.get_num_kv_heads(
                    self.parallel_config),
                head_dim=self.model_config.get_head_size(),
                page_size=self.block_size,
                data_type=kv_cache_dtype)
        else:
            attn_metadata = self.attn_backend.make_metadata(
                is_prompt=False,
                seq_lens=None,
                seq_lens_tensor=seq_lens_tensor,
                max_query_len=None,
                max_seq_len=max_seq_len,
                subquery_start_loc=None,
                seq_start_loc=None,
                context_lens_tensor=None,
                block_tables=block_tables,
                use_cuda_graph=use_captured_graph,
            )
589
590
591
592
593
594
595
596
597
        return PrepareDecodeMetadata(
            input_tokens=input_tokens,
            input_positions=input_positions,
            attn_metadata=attn_metadata,
            lora_index_mapping=lora_index_mapping,
            lora_prompt_mapping=lora_prompt_mapping,
            lora_requests=lora_requests,
            slot_mapping=slot_mapping,
        )
598

599
600
    def prepare_input_tensors(
        self,
601
        seq_group_metadata_list: List[SequenceGroupMetadata],
602
    ) -> Tuple[torch.Tensor, torch.Tensor, AttentionMetadata, SamplingMetadata,
603
               Set[LoRARequest], LoRAMapping, torch.Tensor]:
604
        if self.is_driver_worker:
605
606
607
608
609
610
611
612
            prefill_reqs = []
            decode_reqs = []
            for seq_group_meta in seq_group_metadata_list:
                if seq_group_meta.is_prompt:
                    prefill_reqs.append(seq_group_meta)
                else:
                    decode_reqs.append(seq_group_meta)

613
            # Prepare input tensors.
614
615
616
617
            (
                input_tokens,
                input_positions,
                prefill_attn_metadata,
618
619
                seq_lens,
                query_lens,
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
                lora_index_mapping,
                lora_prompt_mapping,
                lora_requests,
                multi_modal_input,
                slot_mapping,
            ) = self._prepare_prompt(prefill_reqs)
            (
                decode_input_tokens,
                decode_input_positions,
                decode_attn_metadata,
                decode_lora_index_mapping,
                decode_lora_prompt_mapping,
                decode_lora_requests,
                decode_slot_mapping,
            ) = self._prepare_decode(decode_reqs)
635
            sampling_metadata = SamplingMetadata.prepare(
636
637
                seq_group_metadata_list, seq_lens, query_lens, self.device,
                self.pin_memory)
638

639
640
641
            if not self.scheduler_config.chunked_prefill_enabled:
                assert (len(prefill_reqs) and len(decode_reqs)) == 0

642
            num_prefills = len(seq_lens)
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
            num_prefill_tokens = len(input_tokens)
            num_decode_tokens = len(decode_input_tokens)

            # Coalesce tensors. Note that attn_metadata is currently not
            # coalesced for simplicity.
            input_tokens.extend(decode_input_tokens)
            input_positions.extend(decode_input_positions)
            slot_mapping.extend(decode_slot_mapping)
            lora_index_mapping.extend(decode_lora_index_mapping)
            lora_prompt_mapping.extend(decode_lora_prompt_mapping)
            lora_requests.update(decode_lora_requests)

            input_tokens = torch.tensor(input_tokens,
                                        dtype=torch.long,
                                        device=self.device)
            input_positions = torch.tensor(input_positions,
                                           dtype=torch.long,
                                           device=self.device)
            slot_mapping = torch.tensor(slot_mapping,
                                        dtype=torch.long,
                                        device=self.device)

665
666
            if self.lora_config:
                lora_mapping = LoRAMapping(
667
                    lora_index_mapping,
668
669
670
671
672
                    lora_prompt_mapping,
                )
            else:
                lora_mapping = None

673
            # Broadcast the metadata.
674
675
676
677
678
679
680
681
682
683
            # If batch contains both prefill and decode, it sends 2 broadcasts.
            # If it only contains 1 type, it triggers a single broadcast.
            if (prefill_attn_metadata is not None
                    and decode_attn_metadata is not None):
                batch_type = BatchType.MIXED
            elif prefill_attn_metadata is not None:
                batch_type = BatchType.PREFILL
            else:
                batch_type = BatchType.DECODE

684
685
686
687
688
            metadata_dict = {
                "input_tokens": input_tokens,
                "input_positions": input_positions,
                "selected_token_indices":
                sampling_metadata.selected_token_indices,
689
690
                "lora_requests": lora_requests,
                "lora_mapping": lora_mapping,
691
                "multi_modal_input": multi_modal_input,
692
693
694
695
696
                "num_prefill_tokens": num_prefill_tokens,
                "num_decode_tokens": num_decode_tokens,
                "slot_mapping": slot_mapping,
                "num_prefills": num_prefills,
                "batch_type": batch_type,
697
            }
698
699
700
            if prefill_attn_metadata is not None:
                metadata_dict.update(prefill_attn_metadata.asdict_zerocopy())
            else:
701
                assert decode_attn_metadata is not None
702
                metadata_dict.update(decode_attn_metadata.asdict_zerocopy())
703
            broadcast_tensor_dict(metadata_dict, src=0)
704
705
706
707
708
709
710
711

            # Broadcast decode attn metadata for mixed batch type.
            # The additional broadcast costs 300us overhead on 4 A10 GPUs.
            # We can potentially reduce the overhead by coelescing tensors.
            if batch_type == BatchType.MIXED:
                assert decode_attn_metadata is not None
                metadata_dict = decode_attn_metadata.asdict_zerocopy()
                broadcast_tensor_dict(metadata_dict, src=0)
712
        else:
713
            metadata_dict = broadcast_tensor_dict(src=0)
714
715
            input_tokens = metadata_dict.pop("input_tokens")
            input_positions = metadata_dict.pop("input_positions")
716
717
            slot_mapping = metadata_dict.pop("slot_mapping")
            num_prefills = metadata_dict.pop("num_prefills")
718
719
720
721
            selected_token_indices = metadata_dict.pop(
                "selected_token_indices")
            lora_mapping = metadata_dict.pop("lora_mapping")
            lora_requests = metadata_dict.pop("lora_requests")
722
            multi_modal_input = metadata_dict.pop("multi_modal_input")
723
724
725
726
727
728
729
730
731
732
733
734
735
            num_prefill_tokens = metadata_dict.pop("num_prefill_tokens")
            num_decode_tokens = metadata_dict.pop("num_decode_tokens")
            batch_type = metadata_dict.pop("batch_type")

            # Create an attention metadata.
            prefill_attn_metadata = None
            decode_attn_metadata = None
            if batch_type == BatchType.PREFILL or batch_type == BatchType.MIXED:
                prefill_attn_metadata = self.attn_backend.make_metadata(
                    **metadata_dict)
            else:
                decode_attn_metadata = self.attn_backend.make_metadata(
                    **metadata_dict)
736
737
            sampling_metadata = SamplingMetadata(
                seq_groups=None,
738
                selected_token_indices=selected_token_indices,
739
                categorized_sample_indices=None,
740
                num_prompts=0,
741
742
            )

743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
            # if it is a mixed batch, decode attn_metadata is broadcasted
            # separately.
            if batch_type == BatchType.MIXED:
                metadata_dict = broadcast_tensor_dict(src=0)
                decode_attn_metadata = self.attn_backend.make_metadata(
                    **metadata_dict)

        attn_metadata = AttentionMetadata(
            num_prefills=num_prefills,
            slot_mapping=slot_mapping,
            num_prefill_tokens=num_prefill_tokens,
            num_decode_tokens=num_decode_tokens,
            prefill_metadata=prefill_attn_metadata,
            decode_metadata=decode_attn_metadata,
            kv_cache_dtype=self.kv_cache_dtype,
        )

760
        return (input_tokens, input_positions, attn_metadata,
761
762
                sampling_metadata, lora_requests, lora_mapping,
                multi_modal_input)
763

764
765
766
    @torch.inference_mode()
    def execute_model(
        self,
767
        seq_group_metadata_list: List[SequenceGroupMetadata],
768
        kv_caches: List[torch.Tensor],
769
    ) -> Optional[SamplerOutput]:
770
        (input_tokens, input_positions, attn_metadata, sampling_metadata,
771
772
         lora_requests, lora_mapping, multi_modal_input
         ) = self.prepare_input_tensors(seq_group_metadata_list)
773
774
775
776

        if self.lora_config:
            self.set_active_loras(lora_requests, lora_mapping)

777
778
779
780
        # Currently cuda graph is only supported by the decode phase.
        prefill_meta = attn_metadata.prefill_metadata
        decode_meta = attn_metadata.decode_metadata
        if prefill_meta is None and decode_meta.use_cuda_graph:
781
782
783
784
            graph_batch_size = input_tokens.shape[0]
            model_executable = self.graph_runners[graph_batch_size]
        else:
            model_executable = self.model
785
786
787
788
789
790
791
792
793
        execute_model_kwargs = {
            "input_ids": input_tokens,
            "positions": input_positions,
            "kv_caches": kv_caches,
            "attn_metadata": attn_metadata,
        }
        if self.vision_language_config:
            execute_model_kwargs.update({"image_input": multi_modal_input})
        hidden_states = model_executable(**execute_model_kwargs)
794

795
796
797
798
        # Compute the logits.
        logits = self.model.compute_logits(hidden_states, sampling_metadata)

        # Only perform sampling in the driver worker.
799
        if not self.is_driver_worker:
800
801
            return None

802
803
        # Sample the next token.
        output = self.model.sample(
804
            logits=logits,
805
806
            sampling_metadata=sampling_metadata,
        )
807

808
809
810
811
812
        return output

    @torch.inference_mode()
    def profile_run(self) -> None:
        # Enable top-k sampling to reflect the accurate memory usage.
813
        sampling_params = SamplingParams(top_p=0.99, top_k=self.vocab_size - 1)
814
815
816
        max_num_batched_tokens = self.scheduler_config.max_num_batched_tokens
        max_num_seqs = self.scheduler_config.max_num_seqs

817
818
819
820
821
822
823
        # This represents the maximum number of different requests
        # that will have unique loras, an therefore the max amount of memory
        # consumption create dummy lora request copies from the lora request
        # passed in, which contains a lora from the lora warmup path.
        dummy_lora_requests = []
        dummy_lora_requests_per_seq = []
        if self.lora_config:
824
            assert self.lora_manager is not None
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
            with self.lora_manager.dummy_lora_cache():
                for idx in range(self.lora_config.max_loras):
                    lora_id = idx + 1
                    dummy_lora_request = LoRARequest(
                        lora_name=f"warmup_{lora_id}",
                        lora_int_id=lora_id,
                        lora_local_path="/not/a/real/path",
                    )
                    self.lora_manager.add_dummy_lora(dummy_lora_request,
                                                     rank=LORA_WARMUP_RANK)
                    dummy_lora_requests.append(dummy_lora_request)
                dummy_lora_requests_per_seq = [
                    dummy_lora_requests[idx % len(dummy_lora_requests)]
                    for idx in range(max_num_seqs)
                ]
840

841
842
843
        # Profile memory usage with max_num_sequences sequences and the total
        # number of tokens equal to max_num_batched_tokens.
        seqs: List[SequenceGroupMetadata] = []
844
845
846
847
848
849
850
851
852
853
854
        # Additional GPU memory may be needed for vision encoding, which needs
        # to be accounted for when calculating the GPU blocks for
        # vLLM blocker manager.
        # To exercise the worst scenario for GPU memory consumption,
        # the number of seqs (batch_size) is chosen to maximize the number
        # of images processed.
        if self.vision_language_config:
            max_num_seqs = min(
                max_num_seqs,
                int(max_num_batched_tokens /
                    self.vision_language_config.image_feature_size))
855
856
857
        for group_id in range(max_num_seqs):
            seq_len = (max_num_batched_tokens // max_num_seqs +
                       (group_id < max_num_batched_tokens % max_num_seqs))
858
859
            seq_data, fake_multi_modal_input = _prepare_fake_inputs(
                seq_len, self.vision_language_config)
860
861
862
863
864
865
            seq = SequenceGroupMetadata(
                request_id=str(group_id),
                is_prompt=True,
                seq_data={group_id: seq_data},
                sampling_params=sampling_params,
                block_tables=None,
866
867
                lora_request=dummy_lora_requests_per_seq[group_id]
                if dummy_lora_requests_per_seq else None,
868
                multi_modal_data=fake_multi_modal_input,
869
870
871
872
873
            )
            seqs.append(seq)

        # Run the model with the dummy inputs.
        num_layers = self.model_config.get_num_layers(self.parallel_config)
874
        kv_caches = [None] * num_layers
875
        self.execute_model(seqs, kv_caches)
876
        torch.cuda.synchronize()
877
878
        return

879
    def remove_all_loras(self):
880
881
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
882
        self.lora_manager.remove_all_loras()
883

884
    def set_active_loras(self, lora_requests: Set[LoRARequest],
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
                         lora_mapping: LoRAMapping) -> None:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
        self.lora_manager.set_active_loras(lora_requests, lora_mapping)

    def add_lora(self, lora_request: LoRARequest) -> bool:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
        return self.lora_manager.add_lora(lora_request)

    def remove_lora(self, lora_id: int) -> bool:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
        return self.lora_manager.remove_lora(lora_id)

    def list_loras(self) -> Set[int]:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
        return self.lora_manager.list_loras()

905
    @torch.inference_mode()
906
    def capture_model(self, kv_caches: List[torch.Tensor]) -> None:
907
908
909
910
911
912
913
914
915
916
917
918
        """Cuda graph capture a model.

        Note that CUDA graph's performance gain is negligible if number
        of batched tokens are larger than 200. And since CUDA graph
        requires fixed sized tensors, supporting large/variable batch
        size requires high GPU memory overhead. Thus, vLLM only captures
        decoding requests. Mixed batch (chunked prefill + decoding) or
        prefill requests are not captured.

        Since it is used for decoding-only, it assumes there's only 1 token
        per sequence in the batch.
        """
919
920
921
922
923
        assert not self.model_config.enforce_eager
        logger.info("Capturing the model for CUDA graphs. This may lead to "
                    "unexpected consequences if the model is not static. To "
                    "run the model in eager mode, set 'enforce_eager=True' or "
                    "use '--enforce-eager' in the CLI.")
924
925
        logger.info("CUDA graphs can take additional 1~3 GiB memory per GPU. "
                    "If you are running out of memory, consider decreasing "
926
927
928
                    "`gpu_memory_utilization` or enforcing eager mode. "
                    "You can also reduce the `max_num_seqs` as needed "
                    "to decrease memory usage.")
929
930
931
932
        start_time = time.perf_counter()

        # Prepare dummy inputs. These will be reused for all batch sizes.
        max_batch_size = max(_BATCH_SIZES_TO_CAPTURE)
933
934
935
        input_tokens = torch.zeros(max_batch_size, dtype=torch.long).cuda()
        input_positions = torch.zeros(max_batch_size, dtype=torch.long).cuda()
        slot_mapping = torch.empty(max_batch_size, dtype=torch.long).cuda()
936
        slot_mapping.fill_(_PAD_SLOT_ID)
937
        seq_lens = torch.ones(max_batch_size, dtype=torch.int32).cuda()
938
939
        block_tables = torch.from_numpy(self.graph_block_tables).cuda()

940
941
942
943
944
945
        graph_batch_size = _get_graph_batch_size(
            self.scheduler_config.max_num_seqs)
        batch_size_capture_list = [
            bs for bs in _BATCH_SIZES_TO_CAPTURE if bs <= graph_batch_size
        ]

Woosuk Kwon's avatar
Woosuk Kwon committed
946
        # NOTE(woosuk): There are 3 backends for all-reduce: custom all-reduce
947
948
        # kernel, pynccl, and PyTorch NCCL. When using CUDA graph, we use
        # either custom all-reduce kernel or pynccl. When not using CUDA
Woosuk Kwon's avatar
Woosuk Kwon committed
949
950
        # graph, we use either custom all-reduce kernel or PyTorch NCCL.
        # We always prioritize using custom all-reduce kernel but fall back
951
        # to PyTorch or pynccl if it is disabled or not supported.
952
        with custom_all_reduce.capture():
953
954
            # NOTE: Capturing the largest batch size first may help reduce the
            # memory usage of CUDA graph.
955
            for batch_size in reversed(batch_size_capture_list):
956
                # Create dummy attn_metadata.
957
                decode_metadata = self.attn_backend.make_metadata(
958
                    is_prompt=False,
959
960
961
962
                    seq_lens=None,
                    seq_lens_tensor=seq_lens[:batch_size],
                    max_query_len=None,
                    max_seq_len=self.max_seq_len_to_capture,
963
964
                    subquery_start_loc=None,
                    seq_start_loc=None,
965
                    context_lens_tensor=None,
966
967
                    block_tables=block_tables[:batch_size],
                    use_cuda_graph=True,
968
969
970
971
972
973
974
975
                )
                attn_metadata = AttentionMetadata(
                    num_prefills=0,
                    num_prefill_tokens=0,
                    num_decode_tokens=batch_size,
                    slot_mapping=slot_mapping[:batch_size],
                    prefill_metadata=None,
                    decode_metadata=decode_metadata,
976
                    kv_cache_dtype=self.kv_cache_dtype,
977
                )
978

979
980
981
982
983
984
985
986
987
988
989
990
                if self.lora_config:
                    lora_mapping = LoRAMapping(
                        [0] * batch_size,
                        [0] * batch_size,
                    )
                    self.set_active_loras(set(), lora_mapping)

                graph_runner = CUDAGraphRunner(self.model)
                graph_runner.capture(
                    input_tokens[:batch_size],
                    input_positions[:batch_size],
                    kv_caches,
991
                    attn_metadata,
992
                    memory_pool=self.graph_memory_pool,
993
                )
994
995
                self.graph_memory_pool = graph_runner.graph.pool()
                self.graph_runners[batch_size] = graph_runner
996
997
998
999

        end_time = time.perf_counter()
        elapsed_time = end_time - start_time
        # This usually takes < 10 seconds.
1000
        logger.info("Graph capturing finished in %.0f secs.", elapsed_time)
1001

Woosuk Kwon's avatar
Woosuk Kwon committed
1002
    def __del__(self) -> None:
1003
        # Delete the CUDA graphs before deleting the pynccl communicator.
Woosuk Kwon's avatar
Woosuk Kwon committed
1004
1005
1006
        # NOTE(woosuk): This is necessary because otherwise deadlocks can
        # happen.
        # FIXME(woosuk): This is a bit hacky. Find a more robust solution.
1007
1008
        # TODO(youkaichao): when we get enough user feedback that pynccl is
        # more stable than cupy, we can remove this, e.g. in v0.4.1.
Woosuk Kwon's avatar
Woosuk Kwon committed
1009
        self.graph_runners.clear()
1010
        self.pynccl_backend = None
Woosuk Kwon's avatar
Woosuk Kwon committed
1011

1012
1013
1014
1015
    @property
    def vocab_size(self) -> int:
        return self.model_config.get_vocab_size()

1016
1017
1018
1019
1020
1021
1022
1023

class CUDAGraphRunner:

    def __init__(self, model: nn.Module):
        self.model = model
        self.input_buffers: Dict[str, torch.Tensor] = {}
        self.output_buffers: Dict[str, torch.Tensor] = {}

1024
1025
1026
1027
1028
1029
1030
        self._graph: Optional[torch.cuda.CUDAGraph] = None

    @property
    def graph(self):
        assert self._graph is not None
        return self._graph

1031
1032
1033
1034
    def capture(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1035
1036
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
1037
        memory_pool,
1038
        **kwargs,
1039
    ) -> None:
1040
        assert self._graph is None
1041
1042
1043
        # Run the model once without capturing the graph.
        # This is to make sure that the captured graph does not include the
        # kernel launches for initial benchmarking (e.g., Triton autotune).
1044
        with graph_capture_mode():
Woosuk Kwon's avatar
Woosuk Kwon committed
1045
            self.model(
1046
1047
1048
                input_ids,
                positions,
                kv_caches,
1049
                attn_metadata,
1050
                **kwargs,
1051
1052
1053
            )
        torch.cuda.synchronize()

Woosuk Kwon's avatar
Woosuk Kwon committed
1054
1055
1056
        # Capture the graph.
        # NOTE(woosuk): Python 3.8 does not support multi-line with statements.
        # https://stackoverflow.com/questions/31039022/python-multi-line-with-statement
1057
1058
        self._graph = torch.cuda.CUDAGraph()
        with torch.cuda.graph(self._graph, pool=memory_pool):  # noqa: SIM117
1059
            with graph_capture_mode():
Woosuk Kwon's avatar
Woosuk Kwon committed
1060
1061
1062
1063
                hidden_states = self.model(
                    input_ids,
                    positions,
                    kv_caches,
1064
                    attn_metadata,
1065
                    **kwargs,
Woosuk Kwon's avatar
Woosuk Kwon committed
1066
1067
1068
                )
        torch.cuda.synchronize()

1069
1070
1071
1072
1073
        # Save the input and output buffers.
        self.input_buffers = {
            "input_ids": input_ids,
            "positions": positions,
            "kv_caches": kv_caches,
1074
            "slot_mapping": attn_metadata.slot_mapping,
1075
            "seq_lens_tensor": attn_metadata.decode_metadata.seq_lens_tensor,
1076
            "block_tables": attn_metadata.decode_metadata.block_tables,
1077
1078
1079
1080
1081
1082
1083
1084
        }
        self.output_buffers = {"hidden_states": hidden_states}
        return

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1085
1086
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
1087
        **kwargs,
1088
1089
1090
1091
1092
    ) -> torch.Tensor:
        # KV caches are fixed tensors, so we don't need to copy them.
        del kv_caches

        # Copy the input tensors to the input buffers.
1093
1094
        self.input_buffers["input_ids"].copy_(input_ids, non_blocking=True)
        self.input_buffers["positions"].copy_(positions, non_blocking=True)
1095
        self.input_buffers["slot_mapping"].copy_(attn_metadata.slot_mapping,
1096
                                                 non_blocking=True)
1097
1098
        self.input_buffers["seq_lens_tensor"].copy_(
            attn_metadata.decode_metadata.seq_lens_tensor, non_blocking=True)
1099
1100
        self.input_buffers["block_tables"].copy_(
            attn_metadata.decode_metadata.block_tables, non_blocking=True)
1101
1102
1103
1104
1105
1106
1107
1108
1109
        # Run the graph.
        self.graph.replay()

        # Return the output tensor.
        return self.output_buffers["hidden_states"]

    def __call__(self, *args, **kwargs):
        return self.forward(*args, **kwargs)

1110

1111
def _get_graph_batch_size(batch_size: int) -> int:
1112
1113
1114
1115
1116
    """Returns the padded batch size given actual batch size.

    Batch sizes are 1, 2, 4, _BATCH_SIZE_ALIGNMENT,
    2*_BATCH_SIZE_ALIGNMENT, 3*_BATCH_SIZE_ALIGNMENT...
    """
1117
1118
1119
1120
1121
    if batch_size <= 2:
        return batch_size
    elif batch_size <= 4:
        return 4
    else:
1122
1123
        return ((batch_size + _BATCH_SIZE_ALIGNMENT - 1) //
                _BATCH_SIZE_ALIGNMENT * _BATCH_SIZE_ALIGNMENT)
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141


def _prepare_fake_inputs(
        seq_len: int, vision_language_config: Optional[VisionLanguageConfig]):
    """Prepare fake inputs for profile run."""
    if vision_language_config:
        prompt_tokens = [
            vision_language_config.image_token_id
        ] * vision_language_config.image_feature_size + [0] * (
            seq_len - vision_language_config.image_feature_size)
        fake_image_input = MultiModalData(
            type=MultiModalData.Type.IMAGE,
            data=torch.zeros(vision_language_config.image_input_shape,
                             dtype=torch.float16))
    else:
        prompt_tokens = [0] * seq_len
        fake_image_input = None
    return SequenceData(prompt_tokens), fake_image_input