model_runner.py 70.4 KB
Newer Older
1
import dataclasses
2
import gc
3
import time
4
import warnings
5
from collections import defaultdict
6
7
from typing import (TYPE_CHECKING, Any, Dict, List, Mapping, Optional, Set,
                    Tuple, Type, TypeVar, Union)
8

9
import numpy as np
10
import torch
11
import torch.distributed
12
import torch.nn as nn
13

14
15
16
17
try:
    from flashinfer import BatchDecodeWithPagedKVCacheWrapper
    from flashinfer.decode import CUDAGraphBatchDecodeWithPagedKVCacheWrapper
    from flashinfer.prefill import BatchPrefillWithPagedKVCacheWrapper
18
    FLASHINFER_WORKSPACE_BUFFER_SIZE = 256 * 1024 * 1024
19
20
21
22
23
24
except ImportError:
    BatchDecodeWithPagedKVCacheWrapper = None
    CUDAGraphBatchDecodeWithPagedKVCacheWrapper = None
    BatchPrefillWithPagedKVCacheWrapper = None
    FLASHINFER_WORKSPACE_BUFFER_SIZE = 0

25
from vllm.attention import AttentionMetadata, get_attn_backend
26
from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
27
                         ModelConfig, MultiModalConfig, ParallelConfig,
28
                         PromptAdapterConfig, SchedulerConfig)
29
from vllm.distributed import get_pp_group
30
from vllm.distributed.parallel_state import graph_capture
31
from vllm.inputs import INPUT_REGISTRY
32
from vllm.logger import init_logger
33
34
35
from vllm.lora.layers import LoRAMapping
from vllm.lora.request import LoRARequest
from vllm.lora.worker_manager import LRUCacheWorkerLoRAManager
36
from vllm.model_executor import SamplingMetadata
37
from vllm.model_executor.model_loader import get_model
38
from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
39
40
from vllm.model_executor.models.interfaces import (supports_lora,
                                                   supports_vision)
41
42
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensors,
                             MultiModalInputs)
43
44
45
46
from vllm.prompt_adapter.layers import PromptAdapterMapping
from vllm.prompt_adapter.request import PromptAdapterRequest
from vllm.prompt_adapter.worker_manager import (
    LRUCacheWorkerPromptAdapterManager)
47
from vllm.sampling_params import SamplingParams
48
49
from vllm.sequence import (IntermediateTensors, SamplerOutput,
                           SequenceGroupMetadata)
50
51
from vllm.utils import (CudaMemoryProfiler, get_kv_cache_torch_dtype, is_hip,
                        is_pin_memory_available, make_tensor_with_pad)
52
53
54
55
56
57
58
59
60
from vllm.worker.model_runner_base import (
    ModelRunnerBase, ModelRunnerInputBase,
    _add_attn_metadata_broadcastable_dict,
    _add_sampling_metadata_broadcastable_dict,
    _init_attn_metadata_from_tensor_dict,
    _init_sampling_metadata_from_tensor_dict)

if TYPE_CHECKING:
    from vllm.attention.backends.abstract import AttentionBackend
61
62
63
64

logger = init_logger(__name__)

_PAD_SLOT_ID = -1
65
LORA_WARMUP_RANK = 8
66
67
_BATCH_SIZE_ALIGNMENT = 8
# Capture graphs for token size 1, 2, 4, 8, 16, 24, 32, 40, ..., 256.
68
# NOTE: _get_graph_batch_size needs to be updated if this list is changed.
69
70
71
_BATCH_SIZES_TO_CAPTURE = [1, 2, 4] + [
    _BATCH_SIZE_ALIGNMENT * i for i in range(1, 33)
]
72
_NUM_WARMUP_ITERS = 2
73

74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
TModelInputForGPU = TypeVar('TModelInputForGPU', bound="ModelInputForGPU")


@dataclasses.dataclass(frozen=True)
class ModelInputForGPU(ModelRunnerInputBase):
    """
    This base class contains metadata needed for the base model forward pass
    but not metadata for possible additional steps, e.g., sampling. Model
    runners that run additional steps should subclass this method to add
    additional fields.
    """
    input_tokens: Optional[torch.Tensor] = None
    input_positions: Optional[torch.Tensor] = None
    seq_lens: Optional[List[int]] = None
    query_lens: Optional[List[int]] = None
    lora_mapping: Optional["LoRAMapping"] = None
    lora_requests: Optional[Set[LoRARequest]] = None
    attn_metadata: Optional["AttentionMetadata"] = None
92
93
    prompt_adapter_mapping: Optional[PromptAdapterMapping] = None
    prompt_adapter_requests: Optional[Set[PromptAdapterRequest]] = None
94
    multi_modal_kwargs: Optional[Mapping[str, BatchedTensors]] = None
Mor Zusman's avatar
Mor Zusman committed
95
96
    request_ids_to_seq_ids: Optional[Dict[str, List[int]]] = None
    finished_requests_ids: Optional[List[str]] = None
97
    virtual_engine: int = 0
98
99
100
101
102
103
104
105

    def as_broadcastable_tensor_dict(self) -> Dict[str, Any]:
        tensor_dict = {
            "input_tokens": self.input_tokens,
            "input_positions": self.input_positions,
            "lora_requests": self.lora_requests,
            "lora_mapping": self.lora_mapping,
            "multi_modal_kwargs": self.multi_modal_kwargs,
106
107
            "prompt_adapter_mapping": self.prompt_adapter_mapping,
            "prompt_adapter_requests": self.prompt_adapter_requests,
108
            "virtual_engine": self.virtual_engine,
Mor Zusman's avatar
Mor Zusman committed
109
110
            "request_ids_to_seq_ids": self.request_ids_to_seq_ids,
            "finished_requests_ids": self.finished_requests_ids,
111
112
113
        }
        _add_attn_metadata_broadcastable_dict(tensor_dict, self.attn_metadata)
        return tensor_dict
114
115

    @classmethod
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
    def from_broadcasted_tensor_dict(
        cls: Type[TModelInputForGPU],
        tensor_dict: Dict[str, Any],
        attn_backend: Optional["AttentionBackend"] = None,
    ) -> TModelInputForGPU:
        if attn_backend is not None:
            tensor_dict = _init_attn_metadata_from_tensor_dict(
                attn_backend, tensor_dict)
        return cls(**tensor_dict)


@dataclasses.dataclass(frozen=True)
class ModelInputForGPUWithSamplingMetadata(ModelInputForGPU):
    """
    Used by the ModelRunner.
    """
    sampling_metadata: Optional["SamplingMetadata"] = None
    # Used for speculative decoding. We do not broadcast it because it is only
    # used by the driver worker.
    is_prompt: Optional[bool] = None

    def as_broadcastable_tensor_dict(self) -> Dict[str, Any]:
        tensor_dict = {
            "input_tokens": self.input_tokens,
            "input_positions": self.input_positions,
            "lora_requests": self.lora_requests,
            "lora_mapping": self.lora_mapping,
            "multi_modal_kwargs": self.multi_modal_kwargs,
144
145
            "prompt_adapter_mapping": self.prompt_adapter_mapping,
            "prompt_adapter_requests": self.prompt_adapter_requests,
146
            "virtual_engine": self.virtual_engine,
Mor Zusman's avatar
Mor Zusman committed
147
148
            "request_ids_to_seq_ids": self.request_ids_to_seq_ids,
            "finished_requests_ids": self.finished_requests_ids,
149
150
151
152
153
        }
        _add_attn_metadata_broadcastable_dict(tensor_dict, self.attn_metadata)
        _add_sampling_metadata_broadcastable_dict(tensor_dict,
                                                  self.sampling_metadata)
        return tensor_dict
154

155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
    @classmethod
    def from_broadcasted_tensor_dict(
        cls,
        tensor_dict: Dict[str, Any],
        attn_backend: Optional["AttentionBackend"] = None,
    ) -> "ModelInputForGPUWithSamplingMetadata":
        tensor_dict = _init_sampling_metadata_from_tensor_dict(tensor_dict)
        if attn_backend is not None:
            tensor_dict = _init_attn_metadata_from_tensor_dict(
                attn_backend, tensor_dict)
        return cls(**tensor_dict)


class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
    """
    Helper class for shared methods between GPU model runners.
    """
    _model_input_cls: Type[TModelInputForGPU]
173
174
175
176
177
178

    def __init__(
        self,
        model_config: ModelConfig,
        parallel_config: ParallelConfig,
        scheduler_config: SchedulerConfig,
179
        device_config: DeviceConfig,
180
        cache_config: CacheConfig,
181
        load_config: LoadConfig,
182
        lora_config: Optional[LoRAConfig],
183
        kv_cache_dtype: Optional[str] = "auto",
184
        is_driver_worker: bool = False,
185
        prompt_adapter_config: Optional[PromptAdapterConfig] = None,
186
        multimodal_config: Optional[MultiModalConfig] = None,
187
        return_hidden_states: bool = False,
188
189
190
191
    ):
        self.model_config = model_config
        self.parallel_config = parallel_config
        self.scheduler_config = scheduler_config
192
193
        self.device_config = device_config
        self.cache_config = cache_config
194
        self.lora_config = lora_config
195
        self.load_config = load_config
196
        self.is_driver_worker = is_driver_worker
197
        self.prompt_adapter_config = prompt_adapter_config
198
        self.multimodal_config = multimodal_config
199
        self.return_hidden_states = return_hidden_states
200

201
        self.device = self.device_config.device
202
        self.pin_memory = is_pin_memory_available()
203

204
205
206
207
        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
208
209
210
211

        self.graph_runners: List[Dict[int, CUDAGraphRunner]] = [
            {} for _ in range(self.parallel_config.pipeline_parallel_size)
        ]
212
213
        self.graph_memory_pool: Optional[Tuple[
            int, int]] = None  # Set during graph capture.
Mor Zusman's avatar
Mor Zusman committed
214
215
216
217

        self.has_seqlen_agnostic = model_config.contains_seqlen_agnostic_layers(
            parallel_config)

218
        # When using CUDA graph, the input block tables must be padded to
219
        # max_seq_len_to_capture. However, creating the block table in
220
221
222
223
        # 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).
224
225
226
        self.graph_block_tables = np.zeros(
            (max(_BATCH_SIZES_TO_CAPTURE), self.get_max_block_per_batch()),
            dtype=np.int32)
227
228
        num_attn_heads = self.model_config.get_num_attention_heads(
            self.parallel_config)
229
        self.attn_backend = get_attn_backend(
230
            num_attn_heads,
231
232
233
234
235
236
            self.model_config.get_head_size(),
            self.model_config.get_num_kv_heads(self.parallel_config),
            self.model_config.get_sliding_window(),
            self.model_config.dtype,
            self.kv_cache_dtype,
            self.block_size,
237
        ) if num_attn_heads else None
238

239
240
241
        # Multi-modal data support
        self.multi_modal_input_mapper = MULTIMODAL_REGISTRY \
            .create_input_mapper(self.model_config)
242

243
        # Lazy initialization
244
        self.model: nn.Module  # Set after load_model
245
246
        # Set after load_model.
        self.lora_manager: Optional[LRUCacheWorkerLoRAManager] = None
247
        self.prompt_adapter_manager: LRUCacheWorkerPromptAdapterManager = None
248

249
250
251
252
253
        self.flashinfer_decode_workspace_buffer = None
        self.flashinfer_decode_wrapper = None
        self.flashinfer_prefill_workspace_buffer = None
        self.flashinfer_prefill_wrapper = None

254
    def load_model(self) -> None:
255
        with CudaMemoryProfiler() as m:
256
257
258
259
260
261
262
263
            self.model = get_model(model_config=self.model_config,
                                   device_config=self.device_config,
                                   load_config=self.load_config,
                                   lora_config=self.lora_config,
                                   multimodal_config=self.multimodal_config,
                                   parallel_config=self.parallel_config,
                                   scheduler_config=self.scheduler_config,
                                   cache_config=self.cache_config)
264
265

        self.model_memory_usage = m.consumed_memory
266
267
        logger.info("Loading model weights took %.4f GB",
                    self.model_memory_usage / float(2**30))
268
269

        if self.lora_config:
270
            assert supports_lora(self.model), "Model does not support LoRA"
271
272
273
            assert not supports_vision(
                self.model
            ), "To be tested: vision language model with LoRA settings."
274

275
276
            self.lora_manager = LRUCacheWorkerLoRAManager(
                self.scheduler_config.max_num_seqs,
277
278
279
280
281
282
283
284
285
                self.scheduler_config.max_num_batched_tokens,
                self.vocab_size,
                self.lora_config,
                self.device,
                self.model.embedding_modules,
                self.model.embedding_padding_modules,
                max_position_embeddings=self.model.config.
                max_position_embeddings,
            )
286
            self.model = self.lora_manager.create_lora_manager(self.model)
287

288
289
290
291
292
293
294
295
296
        if self.prompt_adapter_config:
            self.prompt_adapter_manager = LRUCacheWorkerPromptAdapterManager(
                self.scheduler_config.max_num_seqs,
                self.scheduler_config.max_num_batched_tokens, self.device,
                self.prompt_adapter_config)
            self.model = (
                self.prompt_adapter_manager.create_prompt_adapter_manager(
                    self.model))

297
        if self.kv_cache_dtype == "fp8" and is_hip():
298
299
300
            # Currently only ROCm accepts kv-cache scaling factors
            # via quantization_param_path and this will be deprecated
            # in the future.
301
302
            if self.model_config.quantization_param_path is not None:
                if callable(getattr(self.model, "load_kv_cache_scales", None)):
303
304
305
306
307
308
                    warnings.warn(
                        "Loading kv cache scaling factor from JSON is "
                        "deprecated and will be removed. Please include "
                        "kv cache scaling factors in the model checkpoint.",
                        FutureWarning,
                        stacklevel=2)
309
310
                    self.model.load_kv_cache_scales(
                        self.model_config.quantization_param_path)
311
312
                    logger.info("Loaded KV cache scaling factors from %s",
                                self.model_config.quantization_param_path)
313
                else:
314
315
316
317
                    raise RuntimeError(
                        "Using FP8 KV cache and scaling factors provided but "
                        "model %s does not support loading scaling factors.",
                        self.model.__class__)
318
            else:
319
320
321
322
                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!")
323

324
325
326
327
328
329
330
331
332
333
334
335
336
337
    def save_sharded_state(
        self,
        path: str,
        pattern: Optional[str] = None,
        max_size: Optional[int] = None,
    ) -> None:
        from vllm.model_executor.model_loader.loader import ShardedStateLoader
        ShardedStateLoader.save_model(
            self.model,
            path,
            pattern=pattern,
            max_size=max_size,
        )

338
339
340
341
342
343
344
345
346
347
    def save_tensorized_model(
        self,
        tensorizer_config: TensorizerConfig,
    ) -> None:
        from vllm.model_executor.model_loader.loader import TensorizerLoader
        TensorizerLoader.save_model(
            self.model,
            tensorizer_config=tensorizer_config,
        )

348
349
    def get_max_block_per_batch(self) -> int:
        block_size = self.block_size
350
        return (self.max_seq_len_to_capture + block_size - 1) // block_size
351

352
    def _prepare_model_input_tensors(
353
354
        self,
        seq_group_metadata_list: List[SequenceGroupMetadata],
Mor Zusman's avatar
Mor Zusman committed
355
        finished_requests_ids: Optional[List[str]] = None
356
357
358
359
    ) -> TModelInputForGPU:
        """Helper method to prepare the model input based on a given sequence
        group. Prepares metadata needed for the base model forward pass but not
        metadata for possible additional steps, e.g., sampling.
360
361
362
363
364
365
366
367
368
369
370

        The API assumes seq_group_metadata_list is sorted by prefill -> decode.

        The result tensors and data structure also batches input in prefill
        -> decode order. For example,

        - input_tokens[:num_prefill_tokens] contains prefill tokens.
        - input_tokens[num_prefill_tokens:] contains decode tokens.

        If cuda graph is required, this API automatically pads inputs.
        """
371
372
373
        input_tokens: List[int] = []
        input_positions: List[int] = []
        slot_mapping: List[int] = []
374
375
376
        lora_index_mapping: List[int] = []
        lora_prompt_mapping: List[int] = []
        lora_requests: Set[LoRARequest] = set()
377
378
379
        prompt_adapter_index_mapping: List[int] = []
        prompt_adapter_prompt_mapping: List[int] = []
        prompt_adapter_requests: Set[PromptAdapterRequest] = set()
380

381
        seq_lens: List[int] = []
382
383
        prefill_seq_lens: List[int] = []
        decode_seq_lens: List[int] = []
384
        context_lens: List[int] = []
385
        query_lens: List[int] = []
386
        block_tables: List[List[int]] = []
387
        multi_modal_inputs_list: List[MultiModalInputs] = []
Mor Zusman's avatar
Mor Zusman committed
388
        request_ids_to_seq_ids: Dict[str, List[int]] = defaultdict(list)
389
390
391
392
        decode_only = True
        num_prefills = 0
        num_prefill_tokens = 0
        num_decode_tokens = 0
393

394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
        # 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] = []

412
        if len(seq_group_metadata_list) == 0:
413
            return self._model_input_cls()
414

415
416
417
418
419
420
        if self.sliding_window is not None:
            sliding_window_blocks = (self.sliding_window + self.block_size -
                                     1) // self.block_size
            block_aligned_sliding_window = \
                sliding_window_blocks * self.block_size

421
422
        for seq_group_metadata in seq_group_metadata_list:
            seq_ids = list(seq_group_metadata.seq_data.keys())
423
            is_prompt = seq_group_metadata.is_prompt
424

425
            for seq_id in seq_ids:
426
427
428
429
430
431
432
433
434
                computed_block_nums = seq_group_metadata.computed_block_nums
                if (self.scheduler_config is not None
                        and self.scheduler_config.chunked_prefill_enabled
                        and not (computed_block_nums is None
                                 or computed_block_nums == [])):
                    raise RuntimeError(
                        "chunked prefill cannot be used with prefix caching "
                        "now.")

435
                seq_data = seq_group_metadata.seq_data[seq_id]
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
                if is_prompt:
                    context_len = seq_data.get_num_computed_tokens()
                else:
                    # get_num_computed_tokens is incorrect for spec decoding.
                    # So, we should have a special logic here.
                    # TODO(sang): Fix it.
                    context_len = seq_data.get_len() - 1

                seq_len = min(
                    seq_data.get_len(),
                    context_len + seq_group_metadata.token_chunk_size)
                if is_prompt:
                    tokens = seq_data.get_token_ids()[context_len:seq_len]
                else:
                    # Optimization. get_token_ids requires the entire copy of
                    # tokens.
                    tokens = [seq_data.get_last_token_id()]
453

454
455
456
457
458
459
460
                # Prefix cache was hit.
                # Prefix is not supported with sliding_window
                prefix_cache_hit = (computed_block_nums is not None
                                    and len(computed_block_nums) > 0
                                    and self.sliding_window is None
                                    and is_prompt)

461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
                # These are seq_len/context_len capped to the sliding window.
                # They are passed to decode kernel.
                # We still need original seq_len/context_len to compute slot
                # mapping (and input position) below.
                curr_sliding_window_blocks = None
                sliding_seq_len = seq_len
                sliding_context_len = context_len

                # TODO(sang): This is a hack to make sliding window work with
                # paged attn. We can remove it if we make paged attn kernel
                # to properly handle slinding window attn.
                if (self.sliding_window is not None and not is_prompt):
                    curr_sliding_window_blocks = sliding_window_blocks
                    if self.scheduler_config.use_v2_block_manager:
                        # number of elements in last block
                        suff_len = seq_len % self.block_size
                        sliding_seq_len = min(
                            seq_len, block_aligned_sliding_window + suff_len)
                        if suff_len > 0:
                            curr_sliding_window_blocks += 1
                    else:
                        sliding_seq_len = min(seq_len, self.sliding_window)
                    sliding_context_len = sliding_seq_len - 1

485
486
487
488
489
490
491
                # 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.
                if prefix_cache_hit:
                    assert computed_block_nums is not None
                    context_len = len(computed_block_nums) * self.block_size
                    tokens = tokens[context_len:]
492
493
494
495
496
497
498

                    # need to think what to set it to when we have both sliding
                    # window and prefix caching...
                    assert self.sliding_window is None, \
                        "Prefix caching is not supported with sliding window"
                    sliding_context_len = context_len

499
500
501
502
503
504
505
506
507
508
509
510
511
                    if self.attn_backend.get_name() == "flash-attn":
                        # NOTE(woosuk): For flash-attn, the block table should
                        # include the entries for the incoming prefill tokens.
                        # TODO(woosuk): This is a temporary fix. We should
                        # provide a unified interface for different backends.
                        block_table = seq_group_metadata.block_tables[seq_id]
                    else:
                        block_table = computed_block_nums
                elif (self.scheduler_config.chunked_prefill_enabled
                      or not is_prompt):
                    if seq_group_metadata.block_tables is not None:
                        # chunked prefill or decode
                        block_table = seq_group_metadata.block_tables[seq_id]
512
513
514
                        if curr_sliding_window_blocks is not None:
                            block_table = block_table[
                                -curr_sliding_window_blocks:]
515
516
517
518
519
520
521
522
                    else:
                        # Only happens when memory profiling runs.
                        block_table = []
                else:
                    # Prefill without chunked prefill or memory profiling.
                    block_table = []
                block_tables.append(block_table)

523
524
525
                seq_lens.append(sliding_seq_len)
                context_lens.append(sliding_context_len)
                query_len = sliding_seq_len - sliding_context_len
526
527
528
529
                query_lens.append(query_len)
                input_tokens.extend(tokens)
                input_positions.extend(list(range(context_len, seq_len)))
                lora_id = seq_group_metadata.lora_int_id
530
                prompt_adapter_id = seq_group_metadata.prompt_adapter_id
531
532
533
534
535
536
537
538
539
540
541
542

                if is_prompt:
                    assert len(seq_ids) == 1
                    num_prefills += 1
                    num_prefill_tokens += len(tokens)
                    decode_only = False
                    prefill_seq_lens.append(seq_len)
                else:
                    assert query_len == 1, (
                        "seq_len: {}, context_len: {}, query_len: {}".format(
                            seq_len, context_len, query_len))
                    num_decode_tokens += query_len
543
                    decode_seq_lens.append(sliding_seq_len)
544
545
546
547

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

548
                lora_index_mapping += [lora_id] * query_len
549
550
                lora_prompt_mapping.extend(
                    [lora_id] *
551
                    (query_len if seq_group_metadata.sampling_params
552
                     and seq_group_metadata.sampling_params.prompt_logprobs
553
                     is not None else 1))
554

555
                mm_data = seq_group_metadata.multi_modal_data
556
                if mm_data:
557
                    # Process multi-modal data
558
                    mm_kwargs = self.multi_modal_input_mapper(mm_data)
559
                    multi_modal_inputs_list.append(mm_kwargs)
560

561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
                if prompt_adapter_id > 0 and is_prompt:
                    prompt_adapter_requests.add(
                        seq_group_metadata.prompt_adapter_request)

                    num_tokens = seq_group_metadata.\
                                            prompt_adapter_num_virtual_tokens
                    pm = [prompt_adapter_id
                          ] * num_tokens + [0] * (query_len - num_tokens)
                    prompt_adapter_index_mapping += pm
                    prompt_adapter_prompt_mapping.extend(
                        [prompt_adapter_id] *
                        (query_len if seq_group_metadata.sampling_params
                         and seq_group_metadata.sampling_params.prompt_logprobs
                         else 1))

576
577
578
                is_profile_run = _is_block_tables_empty(
                    seq_group_metadata.block_tables)
                if is_profile_run:
579
580
581
582
583
584
                    # During memory profiling, the block tables are not
                    # initialized yet. In this case, we just use a dummy
                    # slot mapping.
                    # In embeddings, the block tables are {seq_id: None}.
                    slot_mapping.extend([_PAD_SLOT_ID] * seq_len)
                    continue
585

586
                # Compute the slot mapping.
587
588
                block_table = seq_group_metadata.block_tables[seq_id]

589
590
591
592
593
594
595
                # Mask the [0, start_idx) tokens of the prompt with
                # _PAD_SLOT_ID, where start_idx is max(0, seq_len -
                # sliding_window). 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
596
                if self.sliding_window is not None:
597
                    if is_prompt:
598
599
                        assert self.scheduler_config.use_v2_block_manager \
                            or context_len == 0, (
600
                            "Prefix caching is currently not supported with "
601
                            "sliding window attention in V1 block manager")
602
603
604
605
606
607
608
609
610
611
612
613
614
615
                    # It is an optimization. When it is decoding, it is always
                    # 0. When prefill, we use it to not write slots to kv cache
                    # to save memory.
                    start_idx = max(0, query_len - self.sliding_window)

                for i in range(context_len, seq_len):
                    if i < start_idx:
                        slot_mapping.append(_PAD_SLOT_ID)
                        continue

                    block_number = block_table[i // self.block_size]
                    block_offset = i % self.block_size
                    slot = block_number * self.block_size + block_offset
                    slot_mapping.append(slot)
616

617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
                # Prepare input tensors for flashinfer
                if self.attn_backend.get_name() == "flashinfer":
                    seq_len = seq_data.get_len()
                    # Get the number of valid blocks based on sequence length.
                    # If seq_len = 16, block_size = 16,
                    # block_table_bound is 1 with 1 valid block.
                    # If seq_len = 15, block_size = 16,
                    # block_table_bound is 0 + 1 with 1 valid block.
                    block_table_bound = seq_len // self.block_size + 1 \
                                        if seq_len % self.block_size != 0 \
                                        else seq_len // self.block_size

                    paged_kv_indices.extend(block_table[:block_table_bound])
                    paged_kv_indptr.append(paged_kv_indptr[-1] +
                                           block_table_bound)

                    last_page_len = seq_len % self.block_size
                    if last_page_len == 0:
                        last_page_len = self.block_size
                    paged_kv_last_page_len.append(last_page_len)

638
639
640
641
        batch_size = len(input_tokens)
        max_query_len = max(query_lens)
        max_prefill_seq_len = max(prefill_seq_lens, default=0)
        max_decode_seq_len = max(decode_seq_lens, default=0)
642

643
        # If cuda graph can be used, pad tensors accordingly.
644
        # See `capture_model` API for more details.
645
646
647
648
649
        # vLLM uses cuda graph only for decoding requests.
        use_captured_graph = (
            decode_only and not self.model_config.enforce_eager
            and batch_size <= _BATCH_SIZES_TO_CAPTURE[-1]
            and max_decode_seq_len <= self.max_seq_len_to_capture)
650
651
652
653
        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):
654
655
656
                input_tokens.append(0)
                input_positions.append(0)
                slot_mapping.append(_PAD_SLOT_ID)
657
                seq_lens.append(1)
658
                block_tables.append([])
659
                lora_index_mapping.append(0)
660
                prompt_adapter_index_mapping.append(0)
661
662
663
664
                if self.attn_backend.get_name() == "flashinfer":
                    last_paged_kv_indptr = paged_kv_indptr[-1]
                    paged_kv_indptr.append(last_paged_kv_indptr)
                    paged_kv_last_page_len.append(0)
665
            batch_size = graph_batch_size
666
            num_decode_tokens = batch_size
667
668
669
670
671
672
673
674

        if use_captured_graph:
            # 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
675
            block_tables = torch.tensor(input_block_tables, device=self.device)
676
        else:
677
678
            max_block_table_len = max(
                len(block_table) for block_table in block_tables)
679
            block_tables = make_tensor_with_pad(
680
                block_tables,
681
                max_len=max_block_table_len,
682
683
                pad=0,
                dtype=torch.int,
684
                device=self.device,
685
            )
686
687
        assert max_query_len > 0, ("query_lens: {}".format(query_lens))

688
689
690
691
        context_lens_tensor = torch.tensor(context_lens,
                                           dtype=torch.int,
                                           device=self.device)

692
693
694
        seq_lens_tensor = torch.tensor(seq_lens,
                                       dtype=torch.int,
                                       device=self.device)
695
696
697
698
699
700
        query_lens_tensor = torch.tensor(query_lens,
                                         dtype=torch.long,
                                         device=self.device)
        query_start_loc = torch.zeros(query_lens_tensor.shape[0] + 1,
                                      dtype=torch.int32,
                                      device=self.device)
701
702
703
704
705
706
707
708
        seq_start_loc = torch.zeros(seq_lens_tensor.shape[0] + 1,
                                    dtype=torch.int32,
                                    device=self.device)

        torch.cumsum(seq_lens_tensor,
                     dim=0,
                     dtype=seq_start_loc.dtype,
                     out=seq_start_loc[1:])
709
710
711
712
        torch.cumsum(query_lens_tensor,
                     dim=0,
                     dtype=query_start_loc.dtype,
                     out=query_start_loc[1:])
713
714
715
716
717
718
719
720
721
722

        input_tokens_tensor = torch.tensor(input_tokens,
                                           dtype=torch.long,
                                           device=self.device)
        input_positions_tensor = torch.tensor(input_positions,
                                              dtype=torch.long,
                                              device=self.device)
        slot_mapping_tensor = torch.tensor(slot_mapping,
                                           dtype=torch.long,
                                           device=self.device)
723

724
725
726
727
728
729
730
731
732
733
        logits_soft_cap = getattr(self.model_config.hf_config,
                                  'attn_logit_softcapping', None)
        if logits_soft_cap is not None and self.attn_backend.get_name(
        ) != "flashinfer":
            raise ValueError("Please use Flashinfer backend for models with"
                             "logits_soft_cap (i.e., Gemma-2)."
                             " Otherwise, the output might be wrong."
                             " Set Flashinfer backend by "
                             "export VLLM_ATTENTION_BACKEND=FLASHINFER.")

734
        if self.attn_backend.get_name() == "flashinfer":
735
736
737
738
739
740
741
742
743
744
745
746
747
748
            if len(paged_kv_indptr) > 0:
                paged_kv_indices_tensor = torch.tensor(paged_kv_indices,
                                                       device='cpu',
                                                       dtype=torch.int)
                paged_kv_indptr_tensor = torch.tensor(paged_kv_indptr,
                                                      device='cpu',
                                                      dtype=torch.int)
                paged_kv_last_page_len_tensor = torch.tensor(
                    paged_kv_last_page_len, device='cpu', dtype=torch.int)
            else:
                paged_kv_indices_tensor = None
                paged_kv_indptr_tensor = None
                paged_kv_last_page_len_tensor = None

749
750
751
            kv_cache_dtype = get_kv_cache_torch_dtype(self.kv_cache_dtype,
                                                      self.model_config.dtype)
            attn_metadata = self.attn_backend.make_metadata(
752
753
754
755
756
757
758
759
760
                num_prefills=num_prefills,
                slot_mapping=slot_mapping_tensor,
                num_prefill_tokens=num_prefill_tokens,
                num_decode_tokens=num_decode_tokens,
                max_prefill_seq_len=max_prefill_seq_len,
                block_tables=block_tables,
                paged_kv_indptr=paged_kv_indptr_tensor,
                paged_kv_indices=paged_kv_indices_tensor,
                paged_kv_last_page_len=paged_kv_last_page_len_tensor,
761
762
763
764
765
                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(),
766
                page_size=self.block_size,
767
                seq_start_loc=seq_start_loc,
768
769
770
                query_start_loc=query_start_loc,
                device=self.device,
                data_type=kv_cache_dtype,
771
772
                use_cuda_graph=use_captured_graph,
                logits_soft_cap=logits_soft_cap)
773

774
        else:
775
            attn_metadata = self.attn_backend.make_metadata(
776
777
778
779
780
                num_prefills=num_prefills,
                slot_mapping=slot_mapping_tensor,
                num_prefill_tokens=num_prefill_tokens,
                num_decode_tokens=num_decode_tokens,
                seq_lens=seq_lens,
781
                seq_lens_tensor=seq_lens_tensor,
782
783
784
785
786
787
                max_query_len=max_query_len,
                max_prefill_seq_len=max_prefill_seq_len,
                max_decode_seq_len=max_decode_seq_len,
                query_start_loc=query_start_loc,
                seq_start_loc=seq_start_loc,
                context_lens_tensor=context_lens_tensor,
788
789
790
                block_tables=block_tables,
                use_cuda_graph=use_captured_graph,
            )
791
792
793
794
795
796
797
798
799

        if self.lora_config:
            lora_mapping = LoRAMapping(
                lora_index_mapping,
                lora_prompt_mapping,
            )
        else:
            lora_mapping = None

800
801
802
803
804
805
806
807
        if self.prompt_adapter_config:
            prompt_adapter_mapping = PromptAdapterMapping(
                prompt_adapter_index_mapping,
                prompt_adapter_prompt_mapping,
            )
        else:
            prompt_adapter_mapping = None

808
809
        multi_modal_kwargs = MultiModalInputs.batch(multi_modal_inputs_list,
                                                    device=self.device)
Mor Zusman's avatar
Mor Zusman committed
810
811
812
813
814
        request_ids_to_seq_ids = {
            seq_group_metadata.request_id:
            list(seq_group_metadata.seq_data.keys())
            for seq_group_metadata in seq_group_metadata_list
        }
815
        return self._model_input_cls(
816
817
            input_tokens=input_tokens_tensor,
            input_positions=input_positions_tensor,
818
            attn_metadata=attn_metadata,
819
820
821
            seq_lens=seq_lens,
            query_lens=query_lens,
            lora_mapping=lora_mapping,
822
            lora_requests=lora_requests,
823
            multi_modal_kwargs=multi_modal_kwargs,
Mor Zusman's avatar
Mor Zusman committed
824
            request_ids_to_seq_ids=request_ids_to_seq_ids,
825
826
827
828
            finished_requests_ids=finished_requests_ids,
            prompt_adapter_mapping=prompt_adapter_mapping,
            prompt_adapter_requests=prompt_adapter_requests,
        )
829

830
831
832
    @torch.inference_mode()
    def profile_run(self) -> None:
        # Enable top-k sampling to reflect the accurate memory usage.
833
        sampling_params = SamplingParams(top_p=0.99, top_k=self.vocab_size - 1)
834
835
        max_num_batched_tokens = self.scheduler_config.max_num_batched_tokens
        max_num_seqs = self.scheduler_config.max_num_seqs
836
837
838
839
        # 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.
840
841
        dummy_lora_requests: List[LoRARequest] = []
        dummy_lora_requests_per_seq: List[LoRARequest] = []
842
        if self.lora_config:
843
            assert self.lora_manager is not None
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
            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)
                ]
859

860
861
862
        # Profile memory usage with max_num_sequences sequences and the total
        # number of tokens equal to max_num_batched_tokens.
        seqs: List[SequenceGroupMetadata] = []
863
864
865
866
867
868
        # 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.
869
870
        model_config = self.model_config

871
        if supports_vision(self.model):
872
873
874
875
876
877
878
879
880
881
882
883
884
            max_mm_tokens = MULTIMODAL_REGISTRY \
                .get_max_multimodal_tokens(model_config)
            max_num_seqs_orig = max_num_seqs
            max_num_seqs = min(max_num_seqs,
                               max_num_batched_tokens // max_mm_tokens)
            if max_num_seqs < 1:
                expr = (f"min({max_num_seqs_orig}, "
                        f"{max_num_batched_tokens} // {max_mm_tokens})")
                logger.warning(
                    "Computed max_num_seqs (%s) to be less than 1. "
                    "Setting it to the minimum value of 1.", expr)
                max_num_seqs = 1

885
        batch_size = 0
886
887
888
        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))
889
            batch_size += seq_len
890

891
892
            seq_data, dummy_multi_modal_data = INPUT_REGISTRY \
                .dummy_data_for_profiling(model_config, seq_len)
893
894
895
896
897

            # Having more tokens is over-conservative but otherwise fine
            assert len(seq_data.prompt_token_ids) >= seq_len, (
                f"Expected at least {seq_len} dummy tokens for profiling, "
                f"but got: {len(seq_data.prompt_token_ids)}")
898

899
900
901
902
903
904
            seq = SequenceGroupMetadata(
                request_id=str(group_id),
                is_prompt=True,
                seq_data={group_id: seq_data},
                sampling_params=sampling_params,
                block_tables=None,
905
906
                lora_request=dummy_lora_requests_per_seq[group_id]
                if dummy_lora_requests_per_seq else None,
907
                multi_modal_data=dummy_multi_modal_data,
908
909
910
911
912
            )
            seqs.append(seq)

        # Run the model with the dummy inputs.
        num_layers = self.model_config.get_num_layers(self.parallel_config)
913
        kv_caches = [None] * num_layers
Mor Zusman's avatar
Mor Zusman committed
914
915
916
        finished_requests_ids = [seq.request_id for seq in seqs]
        model_input = self.prepare_model_input(
            seqs, finished_requests_ids=finished_requests_ids)
917
918
919
920
921
922
923
        intermediate_tensors = None
        if not get_pp_group().is_first_rank:
            intermediate_tensors = self.model.make_empty_intermediate_tensors(
                batch_size=batch_size,
                dtype=self.model_config.dtype,
                device=self.device)
        self.execute_model(model_input, kv_caches, intermediate_tensors)
924
        torch.cuda.synchronize()
925
926
        return

927
    def remove_all_loras(self):
928
929
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
930
        self.lora_manager.remove_all_adapters()
931

932
    def set_active_loras(self, lora_requests: Set[LoRARequest],
933
934
935
                         lora_mapping: LoRAMapping) -> None:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
936
        self.lora_manager.set_active_adapters(lora_requests, lora_mapping)
937
938
939
940

    def add_lora(self, lora_request: LoRARequest) -> bool:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
941
        return self.lora_manager.add_adapter(lora_request)
942
943
944
945

    def remove_lora(self, lora_id: int) -> bool:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
946
        return self.lora_manager.remove_adapter(lora_id)
947
948
949
950

    def pin_lora(self, lora_id: int) -> bool:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
951
        return self.lora_manager.pin_adapter(lora_id)
952
953
954
955

    def list_loras(self) -> Set[int]:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
        return self.lora_manager.list_adapters()

    def remove_all_prompt_adapters(self):
        if not self.prompt_adapter_manager:
            raise RuntimeError("PromptAdapter is not enabled.")
        self.prompt_adapter_manager.remove_all_adapters()

    def set_active_prompt_adapters(
            self, prompt_adapter_requests: Set[PromptAdapterRequest],
            prompt_adapter_mapping: PromptAdapterMapping) -> None:
        if not self.prompt_adapter_manager:
            raise RuntimeError("PromptAdapter is not enabled.")
        self.prompt_adapter_manager.set_active_adapters(
            prompt_adapter_requests, prompt_adapter_mapping)

    def add_prompt_adapter(
            self, prompt_adapter_request: PromptAdapterRequest) -> bool:
        if not self.prompt_adapter_manager:
            raise RuntimeError("PromptAdapter is not enabled.")
        return self.prompt_adapter_manager.add_adapter(prompt_adapter_request)

    def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool:
        if not self.prompt_adapter_manager:
            raise RuntimeError("PromptAdapter is not enabled.")
        return self.prompt_adapter_manager.remove_adapter(prompt_adapter_id)

    def pin_prompt_adapter(self, prompt_adapter_id: int) -> bool:
        if not self.prompt_adapter_manager:
            raise RuntimeError("PromptAdapter is not enabled.")
        return self.prompt_adapter_manager.pin_adapter(prompt_adapter_id)

    def list_prompt_adapters(self) -> Set[int]:
        if not self.prompt_adapter_manager:
            raise RuntimeError("PromptAdapter is not enabled.")
        return self.prompt_adapter_manager.list_adapters()
991

992
    @torch.inference_mode()
993
    def capture_model(self, kv_caches: List[List[torch.Tensor]]) -> None:
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
        """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.
        """
1006
1007
1008
1009
1010
        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.")
1011
1012
        logger.info("CUDA graphs can take additional 1~3 GiB memory per GPU. "
                    "If you are running out of memory, consider decreasing "
1013
1014
1015
                    "`gpu_memory_utilization` or enforcing eager mode. "
                    "You can also reduce the `max_num_seqs` as needed "
                    "to decrease memory usage.")
1016
1017
1018
1019
        start_time = time.perf_counter()

        # Prepare dummy inputs. These will be reused for all batch sizes.
        max_batch_size = max(_BATCH_SIZES_TO_CAPTURE)
1020
1021
1022
        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()
1023
        slot_mapping.fill_(_PAD_SLOT_ID)
1024
        seq_lens = torch.ones(max_batch_size, dtype=torch.int32).cuda()
1025
        block_tables = torch.from_numpy(self.graph_block_tables).cuda()
1026
1027
1028
1029
1030
1031
        intermediate_inputs = None
        if not get_pp_group().is_first_rank:
            intermediate_inputs = self.model.make_empty_intermediate_tensors(
                batch_size=max_batch_size,
                dtype=self.model_config.dtype,
                device=self.device)
1032

1033
1034
        # Prepare buffer for outputs. These will be reused for all batch sizes.
        # It will be filled after the first graph capture.
1035
1036
1037
        hidden_or_intermediate_states: List[Optional[torch.Tensor]] = [
            None
        ] * self.parallel_config.pipeline_parallel_size
1038

1039
1040
1041
1042
1043
1044
        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
        ]

1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
        if self.attn_backend.get_name() == "flashinfer":
            # For flashinfer, different batch sizes will share the
            # same workspace buffer.
            decode_workspace_buffer = \
            torch.empty(FLASHINFER_WORKSPACE_BUFFER_SIZE,
                                                dtype=torch.uint8,
                                              device=self.device)
            indices_buffer = torch.empty(max_batch_size *
                                         self.cache_config.num_gpu_blocks,
                                         dtype=torch.int32,
                                         device=self.device)
            indptr_buffer = torch.empty(max_batch_size + 1,
                                        dtype=torch.int32,
                                        device=self.device)
            last_page_len_buffer = torch.empty(max_batch_size,
                                               dtype=torch.int32,
                                               device=self.device)

1063
        with graph_capture() as graph_capture_context:
1064
1065
            # NOTE: Capturing the largest batch size first may help reduce the
            # memory usage of CUDA graph.
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
            for virtual_engine in range(
                    self.parallel_config.pipeline_parallel_size):
                for batch_size in reversed(batch_size_capture_list):
                    if self.attn_backend.get_name() == "flashinfer":
                        indptr_buffer = indptr_buffer[:batch_size + 1]
                        last_page_len_buffer = last_page_len_buffer[:
                                                                    batch_size]

                        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)
                        if num_qo_heads // num_kv_heads >= 4:
                            use_tensor_cores = True
                        else:
                            use_tensor_cores = False
                        decode_wrapper = \
                            CUDAGraphBatchDecodeWithPagedKVCacheWrapper(
                            decode_workspace_buffer, indptr_buffer,
                            indices_buffer, last_page_len_buffer, "NHD",
                            use_tensor_cores)
                        kv_cache_dtype = get_kv_cache_torch_dtype(
                            self.kv_cache_dtype, self.model_config.dtype)

                        paged_kv_indptr_tensor_host = torch.arange(
                            0, batch_size + 1, dtype=torch.int32)
                        paged_kv_indices_tensor_host = torch.arange(
                            0, batch_size, dtype=torch.int32)
                        paged_kv_last_page_len_tensor_host = torch.full(
                            (batch_size, ), self.block_size, dtype=torch.int32)
                        query_start_loc_host = torch.arange(0,
                                                            batch_size + 1,
                                                            dtype=torch.int32)

                        attn_metadata = self.attn_backend.make_metadata(
                            num_prefills=0,
                            slot_mapping=slot_mapping[:batch_size],
                            num_prefill_tokens=0,
                            num_decode_tokens=batch_size,
                            max_prefill_seq_len=0,
                            block_tables=block_tables,
                            paged_kv_indptr=paged_kv_indptr_tensor_host,
                            paged_kv_indices=paged_kv_indices_tensor_host,
                            paged_kv_last_page_len=
                            paged_kv_last_page_len_tensor_host,
                            num_qo_heads=num_qo_heads,
                            num_kv_heads=num_kv_heads,
                            head_dim=self.model_config.get_head_size(),
                            page_size=self.block_size,
                            seq_start_loc=None,
                            query_start_loc=query_start_loc_host,
                            device=self.device,
                            data_type=kv_cache_dtype,
                            use_cuda_graph=True,
                            decode_wrapper=decode_wrapper,
                            prefill_wrapper=None)
                        attn_metadata.begin_forward()
1124
                    else:
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
                        attn_metadata = self.attn_backend.make_metadata(
                            num_prefills=0,
                            num_prefill_tokens=0,
                            num_decode_tokens=batch_size,
                            slot_mapping=slot_mapping[:batch_size],
                            seq_lens=None,
                            seq_lens_tensor=seq_lens[:batch_size],
                            max_query_len=None,
                            max_prefill_seq_len=0,
                            max_decode_seq_len=self.max_seq_len_to_capture,
                            query_start_loc=None,
                            seq_start_loc=None,
                            context_lens_tensor=None,
                            block_tables=block_tables[:batch_size],
                            use_cuda_graph=True,
                        )

                    if self.lora_config:
                        lora_mapping = LoRAMapping(
                            [0] * batch_size,
                            [0] * batch_size,
                        )
                        self.set_active_loras(set(), lora_mapping)

1149
1150
1151
1152
1153
1154
1155
1156
                    if self.prompt_adapter_config:
                        prompt_adapter_mapping = PromptAdapterMapping(
                            [-1] * batch_size,
                            [-1] * batch_size,
                        )
                        self.set_active_prompt_adapters(
                            set(), prompt_adapter_mapping)

1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
                    graph_runner = CUDAGraphRunner(
                        self.model, self.attn_backend.get_name())

                    if self.attn_backend.get_name() == "flashinfer":
                        graph_runner.flashinfer_indptr_buffer = indptr_buffer
                        graph_runner.flashinfer_indices_buffer = indices_buffer
                        graph_runner.flashinfer_last_page_len_buffer = \
                            last_page_len_buffer
                        graph_runner.flashinfer_decode_workspace_buffer = \
                                decode_workspace_buffer
                        graph_runner.flashinfer_decode_wrapper = \
                            decode_wrapper

Mor Zusman's avatar
Mor Zusman committed
1170
1171
                    capture_inputs = {
                        "input_ids":
1172
                        input_tokens[:batch_size],
Mor Zusman's avatar
Mor Zusman committed
1173
                        "positions":
1174
                        input_positions[:batch_size],
Mor Zusman's avatar
Mor Zusman committed
1175
                        "hidden_or_intermediate_states":
1176
1177
1178
1179
1180
                        hidden_or_intermediate_states[
                            virtual_engine]  # type: ignore
                        [:batch_size]
                        if hidden_or_intermediate_states[virtual_engine]
                        is not None else None,
Mor Zusman's avatar
Mor Zusman committed
1181
                        "intermediate_inputs":
1182
1183
                        intermediate_inputs[:batch_size]
                        if intermediate_inputs is not None else None,
Mor Zusman's avatar
Mor Zusman committed
1184
                        "kv_caches":
1185
                        kv_caches[virtual_engine],
Mor Zusman's avatar
Mor Zusman committed
1186
                        "attn_metadata":
1187
                        attn_metadata,
Mor Zusman's avatar
Mor Zusman committed
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
                        "memory_pool":
                        self.graph_memory_pool,
                        "stream":
                        graph_capture_context.stream
                    }
                    if self.has_seqlen_agnostic:
                        # Only used by Mamba-based models CUDA graph atm (Jamba)
                        capture_inputs.update({
                            "seqlen_agnostic_capture_inputs":
                            self.model.get_seqlen_agnostic_capture_inputs(
                                batch_size)
                        })
                    graph_runner.capture(**capture_inputs)
1201
1202
1203
                    self.graph_memory_pool = graph_runner.graph.pool()
                    self.graph_runners[virtual_engine][batch_size] = (
                        graph_runner)
1204
1205
1206
1207

        end_time = time.perf_counter()
        elapsed_time = end_time - start_time
        # This usually takes < 10 seconds.
1208
        logger.info("Graph capturing finished in %.0f secs.", elapsed_time)
1209

1210
1211
1212
1213
    @property
    def vocab_size(self) -> int:
        return self.model_config.get_vocab_size()

1214

1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
    """
    GPU model runner with sampling step.
    """
    _model_input_cls: Type[ModelInputForGPUWithSamplingMetadata] = (
        ModelInputForGPUWithSamplingMetadata)

    def make_model_input_from_broadcasted_tensor_dict(
        self,
        tensor_dict: Dict[str, Any],
    ) -> ModelInputForGPUWithSamplingMetadata:
1226
        model_input = \
1227
1228
1229
            ModelInputForGPUWithSamplingMetadata.from_broadcasted_tensor_dict(
                tensor_dict,
                attn_backend=self.attn_backend,
1230
1231
            )
        return model_input
1232
1233
1234
1235

    def prepare_model_input(
        self,
        seq_group_metadata_list: List[SequenceGroupMetadata],
1236
        virtual_engine: int = 0,
Mor Zusman's avatar
Mor Zusman committed
1237
        finished_requests_ids: Optional[List[str]] = None
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
    ) -> ModelInputForGPUWithSamplingMetadata:
        """Prepare the model input based on a given sequence group, including
        metadata for the sampling step.

        The API assumes seq_group_metadata_list is sorted by prefill -> decode.

        The result tensors and data structure also batches input in prefill
        -> decode order. For example,

        - input_tokens[:num_prefill_tokens] contains prefill tokens.
        - input_tokens[num_prefill_tokens:] contains decode tokens.

        If cuda graph is required, this API automatically pads inputs.
        """
        model_input = self._prepare_model_input_tensors(
Mor Zusman's avatar
Mor Zusman committed
1253
            seq_group_metadata_list, finished_requests_ids)
1254
1255
1256
1257
1258
1259
1260
1261
1262
        sampling_metadata = SamplingMetadata.prepare(seq_group_metadata_list,
                                                     model_input.seq_lens,
                                                     model_input.query_lens,
                                                     self.device,
                                                     self.pin_memory)
        is_prompt = (seq_group_metadata_list[0].is_prompt
                     if seq_group_metadata_list else None)
        return dataclasses.replace(model_input,
                                   sampling_metadata=sampling_metadata,
1263
1264
                                   is_prompt=is_prompt,
                                   virtual_engine=virtual_engine)
1265
1266
1267
1268
1269
1270

    @torch.inference_mode()
    def execute_model(
        self,
        model_input: ModelInputForGPUWithSamplingMetadata,
        kv_caches: List[torch.Tensor],
1271
        intermediate_tensors: Optional[IntermediateTensors] = None,
1272
        num_steps: int = 1,
1273
    ) -> Optional[Union[List[SamplerOutput], IntermediateTensors]]:
1274
1275
1276
        if num_steps > 1:
            raise ValueError("num_steps > 1 is not supported in ModelRunner")

1277
1278
1279
1280
1281
1282
        if self.lora_config:
            assert model_input.lora_requests is not None
            assert model_input.lora_mapping is not None
            self.set_active_loras(model_input.lora_requests,
                                  model_input.lora_mapping)

1283
1284
1285
1286
1287
1288
1289
        if self.prompt_adapter_config:
            assert model_input.prompt_adapter_requests is not None
            assert model_input.prompt_adapter_mapping is not None
            self.set_active_prompt_adapters(
                model_input.prompt_adapter_requests,
                model_input.prompt_adapter_mapping)

1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
        if self.attn_backend.get_name() == "flashinfer":
            assert model_input.attn_metadata is not None
            assert model_input.input_tokens is not None
            if self.flashinfer_decode_workspace_buffer is None:
                self.flashinfer_decode_workspace_buffer = torch.empty(
                    FLASHINFER_WORKSPACE_BUFFER_SIZE,
                    dtype=torch.uint8,
                    device=self.device)
                self.flashinfer_decode_wrapper = \
                    BatchDecodeWithPagedKVCacheWrapper(
                    self.flashinfer_decode_workspace_buffer, "NHD")
                self.flashinfer_prefill_workspace_buffer = torch.empty(
                    FLASHINFER_WORKSPACE_BUFFER_SIZE,
                    dtype=torch.uint8,
                    device=self.device)
                self.flashinfer_prefill_wrapper = \
                    BatchPrefillWithPagedKVCacheWrapper(
                    self.flashinfer_prefill_workspace_buffer, "NHD")

            model_input.attn_metadata.prefill_wrapper = \
                self.flashinfer_prefill_wrapper
            if model_input.attn_metadata.use_cuda_graph:
                batch_size = model_input.input_tokens.shape[0]
                model_input.attn_metadata.decode_wrapper = self.graph_runners[
1314
1315
                    model_input.
                    virtual_engine][batch_size].flashinfer_decode_wrapper
1316
1317
1318
1319
1320
            else:
                model_input.attn_metadata.decode_wrapper = \
                    self.flashinfer_decode_wrapper
            model_input.attn_metadata.begin_forward()

1321
1322
1323
1324
        # Currently cuda graph is only supported by the decode phase.
        assert model_input.attn_metadata is not None
        prefill_meta = model_input.attn_metadata.prefill_metadata
        decode_meta = model_input.attn_metadata.decode_metadata
1325
1326
1327
        # TODO(andoorve): We can remove this once all
        # virtual engines share the same kv cache.
        virtual_engine = model_input.virtual_engine
1328
1329
1330
        if prefill_meta is None and decode_meta.use_cuda_graph:
            assert model_input.input_tokens is not None
            graph_batch_size = model_input.input_tokens.shape[0]
1331
1332
            model_executable = self.graph_runners[virtual_engine][
                graph_batch_size]
1333
1334
1335
1336
        else:
            model_executable = self.model

        multi_modal_kwargs = model_input.multi_modal_kwargs or {}
Mor Zusman's avatar
Mor Zusman committed
1337
1338
1339
1340
        seqlen_agnostic_kwargs = {
            "finished_requests_ids": model_input.finished_requests_ids,
            "request_ids_to_seq_ids": model_input.request_ids_to_seq_ids,
        } if self.has_seqlen_agnostic else {}
1341
        hidden_or_intermediate_states = model_executable(
1342
1343
1344
1345
            input_ids=model_input.input_tokens,
            positions=model_input.input_positions,
            kv_caches=kv_caches,
            attn_metadata=model_input.attn_metadata,
1346
            intermediate_tensors=intermediate_tensors,
1347
            **multi_modal_kwargs,
Mor Zusman's avatar
Mor Zusman committed
1348
            **seqlen_agnostic_kwargs)
1349

1350
1351
1352
1353
1354
        # Compute the logits in the last pipeline stage.
        if not get_pp_group().is_last_rank:
            return hidden_or_intermediate_states

        logits = self.model.compute_logits(hidden_or_intermediate_states,
1355
1356
1357
                                           model_input.sampling_metadata)

        if not self.is_driver_worker:
1358
            return []
1359
1360
1361
1362
1363
1364
1365
1366
1367

        # Sample the next token.
        output: SamplerOutput = self.model.sample(
            logits=logits,
            sampling_metadata=model_input.sampling_metadata,
        )

        if self.return_hidden_states:
            # we only need to pass hidden states of most recent token
1368
1369
            assert model_input.sampling_metadata is not None
            indices = model_input.sampling_metadata.selected_token_indices
1370
            if model_input.is_prompt:
1371
1372
                hidden_states = hidden_or_intermediate_states.index_select(
                    0, indices)
1373
            elif decode_meta.use_cuda_graph:
1374
1375
1376
                hidden_states = hidden_or_intermediate_states[:len(indices)]
            else:
                hidden_states = hidden_or_intermediate_states
1377

1378
1379
            output.hidden_states = hidden_states

1380
        return [output]
1381
1382


1383
1384
class CUDAGraphRunner:

1385
    def __init__(self, model: nn.Module, backend_name: str):
1386
        self.model = model
1387
1388
        self.backend_name = backend_name

1389
1390
1391
        self.input_buffers: Dict[str, torch.Tensor] = {}
        self.output_buffers: Dict[str, torch.Tensor] = {}

1392
1393
        self._graph: Optional[torch.cuda.CUDAGraph] = None

1394
1395
1396
1397
1398
1399
1400
        self.flashinfer_decode_workspace_buffer: Optional[torch.Tensor] = None
        self.flashinfer_indptr_buffer: Optional[torch.Tensor] = None
        self.flashinfer_indices_buffer: Optional[torch.Tensor] = None
        self.flashinfer_last_page_len_buffer: Optional[torch.Tensor] = None
        self.flashinfer_decode_wrapper: Optional[
            CUDAGraphBatchDecodeWithPagedKVCacheWrapper] = None

1401
1402
1403
1404
1405
    @property
    def graph(self):
        assert self._graph is not None
        return self._graph

1406
1407
1408
1409
    def capture(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1410
1411
1412
        hidden_or_intermediate_states: Optional[Union[IntermediateTensors,
                                                      torch.Tensor]],
        intermediate_inputs: Optional[IntermediateTensors],
1413
1414
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
1415
1416
        memory_pool: Optional[Tuple[int, int]],
        stream: torch.cuda.Stream,
1417
        **kwargs,
1418
    ) -> Union[torch.Tensor, IntermediateTensors]:
1419
        assert self._graph is None
1420
        # Run the model a few times without capturing the graph.
1421
1422
        # This is to make sure that the captured graph does not include the
        # kernel launches for initial benchmarking (e.g., Triton autotune).
1423
1424
1425
1426
1427
1428
1429
        # Note one iteration is not enough for torch.jit.script
        for _ in range(_NUM_WARMUP_ITERS):
            self.model(
                input_ids,
                positions,
                kv_caches,
                attn_metadata,
1430
                intermediate_inputs,
1431
1432
                **kwargs,
            )
1433
1434
1435
1436
1437
        torch.cuda.synchronize()

        # Capture the graph.
        self._graph = torch.cuda.CUDAGraph()
        with torch.cuda.graph(self._graph, pool=memory_pool, stream=stream):
1438
            output_hidden_or_intermediate_states = self.model(
1439
1440
1441
                input_ids,
                positions,
                kv_caches,
1442
                attn_metadata,
1443
                intermediate_inputs,
1444
                **kwargs,
1445
            )
1446
1447
1448
1449
1450
1451
1452
1453
            if hidden_or_intermediate_states is not None:
                if get_pp_group().is_last_rank:
                    hidden_or_intermediate_states.copy_(
                        output_hidden_or_intermediate_states)
                else:
                    for key in hidden_or_intermediate_states.tensors:
                        hidden_or_intermediate_states[key].copy_(
                            output_hidden_or_intermediate_states[key])
1454
            else:
1455
1456
1457
1458
                hidden_or_intermediate_states = (
                    output_hidden_or_intermediate_states)

            del output_hidden_or_intermediate_states
1459
1460
1461
            # make sure `output_hidden_states` is deleted
            # in the graph's memory pool
            gc.collect()
1462
1463
1464
        torch.cuda.synchronize()

        # Save the input and output buffers.
1465
1466
1467
1468
1469
1470
        if self.backend_name == "flashinfer":
            self.input_buffers = {
                "input_ids": input_ids,
                "positions": positions,
                "kv_caches": kv_caches,
                "slot_mapping": attn_metadata.slot_mapping,
Mor Zusman's avatar
Mor Zusman committed
1471
                **kwargs,
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
            }
        else:
            self.input_buffers = {
                "input_ids": input_ids,
                "positions": positions,
                "kv_caches": kv_caches,
                "slot_mapping": attn_metadata.slot_mapping,
                "seq_lens_tensor":
                attn_metadata.decode_metadata.seq_lens_tensor,
                "block_tables": attn_metadata.decode_metadata.block_tables,
Mor Zusman's avatar
Mor Zusman committed
1482
                **kwargs,
1483
            }
1484
1485
1486
1487
1488
1489
1490
1491
1492
        if intermediate_inputs is not None:
            self.input_buffers.update(intermediate_inputs.tensors)
        if get_pp_group().is_last_rank:
            self.output_buffers = {
                "hidden_states": hidden_or_intermediate_states
            }
        else:
            self.output_buffers = hidden_or_intermediate_states
        return hidden_or_intermediate_states
1493
1494
1495
1496
1497

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1498
1499
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
1500
        intermediate_tensors: Optional[IntermediateTensors],
1501
        **kwargs,
1502
1503
1504
1505
1506
    ) -> 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.
1507
1508
        self.input_buffers["input_ids"].copy_(input_ids, non_blocking=True)
        self.input_buffers["positions"].copy_(positions, non_blocking=True)
1509
        self.input_buffers["slot_mapping"].copy_(attn_metadata.slot_mapping,
1510
                                                 non_blocking=True)
1511
1512
1513
1514
1515
1516
        if self.backend_name != "flashinfer":
            self.input_buffers["seq_lens_tensor"].copy_(
                attn_metadata.decode_metadata.seq_lens_tensor,
                non_blocking=True)
            self.input_buffers["block_tables"].copy_(
                attn_metadata.decode_metadata.block_tables, non_blocking=True)
Mor Zusman's avatar
Mor Zusman committed
1517
1518
1519
        if "seqlen_agnostic_capture_inputs" in self.input_buffers:
            self.model.copy_inputs_before_cuda_graphs(self.input_buffers,
                                                      **kwargs)
1520
1521
1522
1523
        if intermediate_tensors is not None:
            for key in intermediate_tensors.tensors:
                self.input_buffers[key].copy_(intermediate_tensors[key],
                                              non_blocking=True)
1524
1525
        # Run the graph.
        self.graph.replay()
Mor Zusman's avatar
Mor Zusman committed
1526
1527
1528
        if "seqlen_agnostic_capture_inputs" in self.input_buffers:
            self.model.copy_outputs_after_cuda_graphs(self.input_buffers,
                                                      **kwargs)
1529
        # Return the output tensor.
1530
1531
1532
1533
        if get_pp_group().is_last_rank:
            return self.output_buffers["hidden_states"]

        return self.output_buffers
1534
1535
1536
1537

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

1538

1539
def _get_graph_batch_size(batch_size: int) -> int:
1540
1541
1542
1543
1544
    """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...
    """
1545
1546
1547
1548
1549
    if batch_size <= 2:
        return batch_size
    elif batch_size <= 4:
        return 4
    else:
1550
1551
        return ((batch_size + _BATCH_SIZE_ALIGNMENT - 1) //
                _BATCH_SIZE_ALIGNMENT * _BATCH_SIZE_ALIGNMENT)
1552
1553


1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
def _is_block_tables_empty(block_tables: Union[None, Dict]):
    """
    Check if block_tables is None or a dictionary with all None values.
    """
    if block_tables is None:
        return True
    if isinstance(block_tables, dict) and all(
            value is None for value in block_tables.values()):
        return True
    return False