model_runner.py 86.5 KB
Newer Older
1
import dataclasses
2
import gc
3
import inspect
4
import itertools
5
import time
6
import warnings
7
import weakref
8
from dataclasses import dataclass
9
10
from typing import (TYPE_CHECKING, Any, Callable, Dict, List, Optional, Set,
                    Tuple, Type, TypeVar, Union)
11

12
import numpy as np
13
import torch
14
import torch.distributed
15
import torch.nn as nn
16

17
import vllm.envs as envs
18
from vllm.attention import AttentionMetadata, get_attn_backend
19
from vllm.attention.backends.abstract import AttentionState
20
from vllm.attention.backends.utils import CommonAttentionState
21
from vllm.config import CompilationLevel, VllmConfig
22
from vllm.core.scheduler import SchedulerOutputs
23
from vllm.distributed import get_kv_transfer_group, get_pp_group
24
from vllm.distributed.parallel_state import graph_capture
25
from vllm.forward_context import set_forward_context
26
from vllm.inputs import INPUT_REGISTRY, InputRegistry
27
from vllm.logger import init_logger
28
29
30
from vllm.lora.layers import LoRAMapping
from vllm.lora.request import LoRARequest
from vllm.lora.worker_manager import LRUCacheWorkerLoRAManager
31
from vllm.model_executor import SamplingMetadata, SamplingMetadataCache
32
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
33
from vllm.model_executor.layers.sampler import SamplerOutput
34
from vllm.model_executor.model_loader import get_model
35
from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
36
from vllm.model_executor.models import supports_lora, supports_multimodal
37
from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
38
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
39
                             MultiModalKwargs, MultiModalPlaceholderMap,
40
                             MultiModalRegistry)
41
from vllm.platforms import current_platform
42
43
44
45
from vllm.prompt_adapter.layers import PromptAdapterMapping
from vllm.prompt_adapter.request import PromptAdapterRequest
from vllm.prompt_adapter.worker_manager import (
    LRUCacheWorkerPromptAdapterManager)
46
from vllm.sampling_params import SamplingParams
47
from vllm.sequence import IntermediateTensors, SequenceGroupMetadata
48
49
50
51
from vllm.utils import (DeviceMemoryProfiler, GiB_bytes, PyObjectCache,
                        async_tensor_h2d, flatten_2d_lists,
                        is_pin_memory_available, supports_dynamo,
                        weak_ref_tensor)
52
from vllm.worker.model_runner_base import (
53
    ModelRunnerBase, ModelRunnerInputBase, ModelRunnerInputBuilderBase,
54
55
56
    _add_attn_metadata_broadcastable_dict,
    _add_sampling_metadata_broadcastable_dict,
    _init_attn_metadata_from_tensor_dict,
57
    _init_sampling_metadata_from_tensor_dict, dump_input_when_exception)
58
59
60

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

logger = init_logger(__name__)

64
LORA_WARMUP_RANK = 8
65

66
_NUM_WARMUP_ITERS = 2
67

68
69
TModelInputForGPU = TypeVar('TModelInputForGPU', bound="ModelInputForGPU")

70
71
72
73
# For now, bump up cache limits for recompilations during CUDA graph warmups.
torch._dynamo.config.cache_size_limit = 128
torch._dynamo.config.accumulated_cache_size_limit = 128

74

75
@dataclass(frozen=True)
76
77
78
79
80
81
82
83
84
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
85
    token_types: Optional[torch.Tensor] = None
86
87
88
89
90
    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
91
92
    prompt_adapter_mapping: Optional[PromptAdapterMapping] = None
    prompt_adapter_requests: Optional[Set[PromptAdapterRequest]] = None
93
    multi_modal_kwargs: Optional[BatchedTensorInputs] = None
Mor Zusman's avatar
Mor Zusman committed
94
95
    request_ids_to_seq_ids: Optional[Dict[str, List[int]]] = None
    finished_requests_ids: Optional[List[str]] = None
96
    virtual_engine: int = 0
97
    async_callback: Optional[Callable] = None
98
99
    seq_group_metadata_list: Optional[List[SequenceGroupMetadata]] = None
    scheduler_outputs: Optional[SchedulerOutputs] = None
100
101
102
103
104
105
106
107

    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,
108
109
            "prompt_adapter_mapping": self.prompt_adapter_mapping,
            "prompt_adapter_requests": self.prompt_adapter_requests,
110
            "virtual_engine": self.virtual_engine,
Mor Zusman's avatar
Mor Zusman committed
111
112
            "request_ids_to_seq_ids": self.request_ids_to_seq_ids,
            "finished_requests_ids": self.finished_requests_ids,
113
114
115
        }
        _add_attn_metadata_broadcastable_dict(tensor_dict, self.attn_metadata)
        return tensor_dict
116
117

    @classmethod
118
119
120
121
122
123
124
125
126
127
    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)

128
129
130
131
132
133
134
135
136
137
138
139
    # Exclude `async_callback` to be able to pickle this object
    def __getstate__(self):
        state = self.__dict__.copy()
        del state["async_callback"]
        return state

    # TODO: What happens when we depickle this object?
    # How can we update this callback to properly pass it to the engine?
    def __setstate__(self, state):
        self.__dict__.update(state)
        self.__dict__.update({'async_callback': None})

140

141
@dataclass(frozen=True)
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
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,
158
159
            "prompt_adapter_mapping": self.prompt_adapter_mapping,
            "prompt_adapter_requests": self.prompt_adapter_requests,
160
            "virtual_engine": self.virtual_engine,
Mor Zusman's avatar
Mor Zusman committed
161
162
            "request_ids_to_seq_ids": self.request_ids_to_seq_ids,
            "finished_requests_ids": self.finished_requests_ids,
163
164
165
166
167
        }
        _add_attn_metadata_broadcastable_dict(tensor_dict, self.attn_metadata)
        _add_sampling_metadata_broadcastable_dict(tensor_dict,
                                                  self.sampling_metadata)
        return tensor_dict
168

169
170
171
172
173
174
175
176
177
178
179
180
181
    @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)


182
class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
183
184
    """Build ModelInputForGPU from SequenceGroupMetadata."""

185
186
187
    # Note: ideally we would be using a dataclass(kw_only=True)
    # here, so that this can be subclassed easily,
    # but kw_only is not supported in python<3.10.
188
189
    class InterDataForSeqGroup:
        """Intermediate data for the current sequence group."""
190

191
192
193
        def simple_reinit(self):
            self.input_tokens[0].clear()  # type: ignore
            self.input_positions[0].clear()  # type: ignore
194
            self.token_types[0].clear()  # type: ignore
195
            self.mrope_input_positions = None  # type: ignore
196
197
198
199
200
201
202
203
204
205
206
            self.seq_lens[0] = 0  # type: ignore
            self.orig_seq_lens[0] = 0  # type: ignore
            self.query_lens[0] = 0  # type: ignore
            self.context_lens[0] = 0  # type: ignore
            self.curr_sliding_window_blocks[0] = 0  # type: ignore
            self.lora_index_mapping.clear()  # type: ignore
            self.lora_prompt_mapping.clear()  # type: ignore
            self.lora_requests.clear()  # type: ignore
            self.prompt_adapter_index_mapping.clear()  # type: ignore
            self.prompt_adapter_prompt_mapping.clear()  # type: ignore

207
208
209
210
211
212
213
214
215
216
217
218
219
220
        def __init__(
            self,
            *,
            # From sequence group metadata.
            request_id: str,
            seq_ids: List[int],
            is_prompt: bool,
            block_tables: Optional[Dict[int, List[int]]],
            computed_block_nums: List[int],
            n_seqs: int = 0,

            # Input tokens and positions.
            input_tokens: Optional[List[List[int]]] = None,
            input_positions: Optional[List[List[int]]] = None,
221
            token_types: Optional[List[List[int]]] = None,
222
            mrope_input_positions: Optional[List[List[List[int]]]] = None,
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246

            # The sequence length (may be capped to the sliding window).
            seq_lens: Optional[List[int]] = None,
            # The original sequence length (before applying sliding window).
            # This is used to compute slot mapping.
            orig_seq_lens: Optional[List[int]] = None,
            # The query length.
            query_lens: Optional[List[int]] = None,
            # The number of tokens that are already computed.
            context_lens: Optional[List[int]] = None,
            # The current sliding window block.
            curr_sliding_window_blocks: Optional[List[int]] = None,

            # LoRA inputs.
            lora_index_mapping: Optional[List[List[int]]] = None,
            lora_prompt_mapping: Optional[List[List[int]]] = None,
            lora_requests: Optional[Set[LoRARequest]] = None,

            # Prompt adapter inputs.
            prompt_adapter_index_mapping: Optional[List[int]] = None,
            prompt_adapter_prompt_mapping: Optional[List[int]] = None,
            prompt_adapter_request: Optional[PromptAdapterRequest] = None,

            # Multi-modal inputs.
247
            multi_modal_kwargs: Optional[MultiModalKwargs] = None,
248
249
            multi_modal_placeholder_maps: Optional[Dict[
                str, MultiModalPlaceholderMap]] = None,
250
251
252

            # Whether the prefix cache is hit (prefill only).
            prefix_cache_hit: bool = False,
253
254
            reinit: bool = False,
            reinit_use_defaults: bool = False,
255
            encoder_seq_len: int = 0,
256
        ):
257
258
259
260
261
262
263
            if reinit:
                assert len(self.seq_ids) == len(seq_ids)  # type: ignore
                for i, seq_id in enumerate(seq_ids):
                    self.seq_ids[i] = seq_id  # type: ignore
            else:
                self.seq_ids = seq_ids

264
265
266
267
268
            self.request_id = request_id
            self.is_prompt = is_prompt
            self.block_tables = block_tables
            self.computed_block_nums = computed_block_nums
            self.n_seqs = n_seqs
269
            self.encoder_seq_len = encoder_seq_len
270

271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
            if reinit:
                if len(self.seq_ids) == 1 and reinit_use_defaults:
                    self.simple_reinit()
                else:
                    if input_tokens:
                        self.input_tokens = input_tokens
                    else:
                        for seq_id in range(len(self.seq_ids)):
                            self.input_tokens[seq_id].clear()

                    if input_positions:
                        self.input_positions = input_positions
                    else:
                        for seq_id in range(len(self.seq_ids)):
                            self.input_positions[seq_id].clear()

287
288
289
290
291
292
                    if token_types:
                        self.token_types = token_types
                    else:
                        for seq_id in range(len(self.seq_ids)):
                            self.token_types[seq_id].clear()

293
294
                    self.mrope_input_positions = None

295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
                    if seq_lens:
                        self.seq_lens = seq_lens
                    else:
                        for seq_id in range(len(self.seq_ids)):
                            self.seq_lens[seq_id] = 0

                    if orig_seq_lens:
                        self.orig_seq_lens = orig_seq_lens
                    else:
                        for seq_id in range(len(self.seq_ids)):
                            self.orig_seq_lens[seq_id] = 0

                    if query_lens:
                        self.query_lens = query_lens
                    else:
                        for seq_id in range(len(self.seq_ids)):
                            self.query_lens[seq_id] = 0

                    if context_lens:
                        self.context_lens = context_lens
                    else:
                        for seq_id in range(len(self.seq_ids)):
                            self.context_lens[seq_id] = 0

                    if curr_sliding_window_blocks:
                        self.curr_sliding_window_blocks = \
                            curr_sliding_window_blocks
                    else:
                        for seq_id in range(len(self.seq_ids)):
                            self.curr_sliding_window_blocks[seq_id] = 0

                    if lora_index_mapping:
                        self.lora_index_mapping = lora_index_mapping
                    else:
                        self.lora_index_mapping.clear()

                    if lora_prompt_mapping:
                        self.lora_prompt_mapping = lora_prompt_mapping
                    else:
                        self.lora_prompt_mapping.clear()

                    if lora_requests:
                        self.lora_requests = lora_requests
                    else:
                        self.lora_requests.clear()

                    if prompt_adapter_index_mapping:
                        self.prompt_adapter_index_mapping = \
                            prompt_adapter_index_mapping
                    else:
                        self.prompt_adapter_index_mapping.clear()

                    if prompt_adapter_prompt_mapping:
                        self.prompt_adapter_prompt_mapping = \
                            prompt_adapter_prompt_mapping
                    else:
                        self.prompt_adapter_prompt_mapping.clear()

            else:
                self.input_tokens = input_tokens or []
                self.input_positions = input_positions or []
356
                self.token_types = token_types or []
357
                self.mrope_input_positions = mrope_input_positions or None
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
                self.seq_lens = seq_lens or []
                self.orig_seq_lens = orig_seq_lens or []
                self.query_lens = query_lens or []
                self.context_lens = context_lens or []
                self.curr_sliding_window_blocks = \
                    curr_sliding_window_blocks or []

                self.lora_index_mapping = lora_index_mapping or []
                self.lora_prompt_mapping = lora_prompt_mapping or []
                self.lora_requests = lora_requests or set()

                self.prompt_adapter_index_mapping = (
                    prompt_adapter_index_mapping or [])
                self.prompt_adapter_prompt_mapping = (
                    prompt_adapter_prompt_mapping or [])

            self.prompt_adapter_request = prompt_adapter_request
375
            self.multi_modal_kwargs = multi_modal_kwargs
376
            self.multi_modal_placeholder_maps = multi_modal_placeholder_maps
377
378
            self.prefix_cache_hit = prefix_cache_hit

379
380
            self.n_seqs = len(self.seq_ids)

381
382
            if not reinit:
                self.__post_init__()
383
384
385
386
387
388

        def __post_init__(self):
            self.n_seqs = len(self.seq_ids)

            self.input_tokens = [[] for _ in range(self.n_seqs)]
            self.input_positions = [[] for _ in range(self.n_seqs)]
389
            self.token_types = [[] for _ in range(self.n_seqs)]
390
            self.mrope_input_positions = None
391
392
393
394
395
396
            self.seq_lens = [0] * self.n_seqs
            self.orig_seq_lens = [0] * self.n_seqs
            self.query_lens = [0] * self.n_seqs
            self.context_lens = [0] * self.n_seqs
            self.curr_sliding_window_blocks = [0] * self.n_seqs

397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
            self.lora_index_mapping = []
            self.lora_prompt_mapping = []

    def gen_inter_data_builder(self, num_seqs: int):
        return lambda: ModelInputForGPUBuilder.InterDataForSeqGroup(
            request_id="",
            seq_ids=[0] * num_seqs,
            is_prompt=True,
            block_tables=None,
            computed_block_nums=[])

    def init_cached_inter_data(self, *args, **kwargs):
        assert len(args) == 0
        assert "seq_ids" in kwargs
        seq_ids = kwargs["seq_ids"]
        num_seqs = len(seq_ids)

        # The inter-data cache is per model_runner
        inter_data_cache = self.runner.inter_data_cache
        if num_seqs not in inter_data_cache:
            inter_data_cache[num_seqs] = PyObjectCache(
                self.gen_inter_data_builder(num_seqs))

        obj = inter_data_cache[num_seqs].get_object()
        obj.__init__(*args, **kwargs)
        return obj

    def reset_cached_inter_data(self):
        for cache in self.runner.inter_data_cache.values():
            cache.reset()
427
428
429
430
431

    def __init__(self,
                 runner: "GPUModelRunnerBase",
                 finished_requests_ids: Optional[List[str]] = None):
        super().__init__()
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
        # Compute functions for each sequence in a sequence group.
        # WARNING: The order of the functions matters!
        self.per_seq_compute_fns = [
            self._compute_lens,
            self._compute_for_prefix_cache_hit,
            self._compute_for_sliding_window,
            self._compute_lora_input,
        ]
        # Compute functions for each sequence group.
        # WARNING: The order of the functions matters!
        self.per_seq_group_compute_fns = [
            self._compute_prompt_adapter_input,
            self._compute_multi_modal_input,
        ]

447
448
449
450
451
452
453
454
455
456
457
458
459
        self.runner = runner
        self.model_input_cls = self.runner._model_input_cls
        self.attn_backend = self.runner.attn_backend
        self.scheduler_config = self.runner.scheduler_config
        self.sliding_window = self.runner.sliding_window
        self.block_size = self.runner.block_size
        self.enable_lora = self.runner.lora_config is not None
        self.enable_prompt_adapter = (self.runner.prompt_adapter_config
                                      is not None)
        self.multi_modal_input_mapper = self.runner.multi_modal_input_mapper
        self.finished_requests_ids = finished_requests_ids
        self.decode_only = True

460
461
462
463
        # Intermediate data (data in CPU before going to GPU) for
        # the current sequence group.
        self.inter_data_list: List[
            ModelInputForGPUBuilder.InterDataForSeqGroup] = []
464
465
466

        # Attention metadata inputs.
        self.attn_metadata_builder = self.attn_backend.make_metadata_builder(
467
            weakref.proxy(self))
468
469
470
471
472
473
474
475
476
477
478

        # Engine/Model configurations.
        self.chunked_prefill_enabled = (
            self.scheduler_config is not None
            and self.scheduler_config.chunked_prefill_enabled)
        if self.sliding_window is not None:
            self.sliding_window_blocks = (
                self.sliding_window + self.block_size - 1) // self.block_size
            self.block_aligned_sliding_window = \
                self.sliding_window_blocks * self.block_size

479
480
481
482
483
484
485
    def _compute_lens(self, inter_data: InterDataForSeqGroup, seq_idx: int,
                      seq_group_metadata: SequenceGroupMetadata):
        """Compute context length, sequence length and tokens
        for the given sequence data.
        """
        seq_data = seq_group_metadata.seq_data[inter_data.seq_ids[seq_idx]]
        token_chunk_size = seq_group_metadata.token_chunk_size
486

487
488
        # Compute context length (the number of tokens that are
        # already computed) and sequence length (total number of tokens).
489

490
491
492
        seq_len = seq_data.get_len()
        if inter_data.is_prompt:
            context_len = seq_data.get_num_computed_tokens()
493
494
            seq_len = min(seq_len, context_len + token_chunk_size)
        elif self.runner.scheduler_config.is_multi_step or \
495
            self.runner.model_config.is_encoder_decoder:
496
            context_len = seq_len - 1
497
498
        else:
            context_len = seq_data.get_num_computed_tokens()
499
500

        # Compute tokens.
501
        tokens = seq_data.get_token_ids()[context_len:seq_len]
502
        token_types = seq_group_metadata.token_type_ids
503
504
505
506

        inter_data.seq_lens[seq_idx] = seq_len
        inter_data.orig_seq_lens[seq_idx] = seq_len
        inter_data.context_lens[seq_idx] = context_len
507
508
        inter_data.input_tokens[seq_idx].extend(tokens)
        inter_data.input_positions[seq_idx].extend(range(context_len, seq_len))
509
510
        inter_data.token_types[seq_idx].extend(
            token_types if token_types else [])
511
        inter_data.query_lens[seq_idx] = seq_len - context_len
512

513
514
515
516
517
518
519
520
521
522
523
        if seq_data.mrope_position_delta is not None:
            if inter_data.mrope_input_positions is None:
                inter_data.mrope_input_positions = [None] * inter_data.n_seqs

            inter_data.mrope_input_positions[
                seq_idx] = MRotaryEmbedding.get_next_input_positions(
                    seq_data.mrope_position_delta,
                    context_len,
                    seq_len,
                )

524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
    def _compute_for_prefix_cache_hit(
            self, inter_data: InterDataForSeqGroup, seq_idx: int,
            seq_group_metadata: SequenceGroupMetadata):
        """Check if hit prefix cache (i.e., some blocks are already computed).
        If hit, update input tokens and positions to only compute the
        remaining blocks.
        """
        computed_block_nums = inter_data.computed_block_nums

        # Note that prefix caching does not support sliding window.
        prefix_cache_hit = (computed_block_nums is not None
                            and len(computed_block_nums) > 0
                            and self.sliding_window is None
                            and inter_data.is_prompt)
        inter_data.prefix_cache_hit = prefix_cache_hit
539
540
541
542
543
544
545
546
547

        if not prefix_cache_hit:
            return

        assert computed_block_nums is not None
        # The cache hit prompt tokens in this sequence. Note that
        # this may be larger than the sequence length if chunked
        # prefill is enabled.
        prefix_cache_len = len(computed_block_nums) * self.block_size
548
549
550
        seq_group_metadata.seq_data[inter_data.seq_ids[
            seq_idx]].update_num_cached_tokens(prefix_cache_len)

551
552
553
554
555
556
557
558
559
560
561
562
563
        # The number of so far computed prompt tokens in this sequence.
        context_len = inter_data.context_lens[seq_idx]
        # The total number of prompt tokens in this sequence.
        # When chunked prefill is enabled, this is the token number of
        # computed chunks + current chunk.
        seq_len = inter_data.seq_lens[seq_idx]
        if prefix_cache_len <= context_len:
            # We already passed the cache hit region,
            # so do normal computation.
            pass
        elif context_len < prefix_cache_len < seq_len:
            # Partial hit. Compute the missing part.
            uncomputed_start = prefix_cache_len - context_len
564
            inter_data.input_tokens[seq_idx] = inter_data.input_tokens[
565
                seq_idx][uncomputed_start:]
566
            inter_data.input_positions[seq_idx] = inter_data.input_positions[
567
                seq_idx][uncomputed_start:]
568
569
            inter_data.token_types[seq_idx] = inter_data.token_types[seq_idx][
                uncomputed_start:]
570
571
            context_len = prefix_cache_len

572
573
574
            inter_data.context_lens[seq_idx] = context_len
            inter_data.query_lens[
                seq_idx] = inter_data.seq_lens[seq_idx] - context_len
575
576
577
578
579
580
581
582
583
        elif seq_len <= prefix_cache_len:
            # Full hit. Only compute the last token to avoid
            # erroneous behavior. FIXME: Ideally we should directly
            # mark all tokens as computed in the scheduler and do not
            # schedule this sequence, so this case should not happen.
            inter_data.input_tokens[seq_idx] = inter_data.input_tokens[
                seq_idx][-1:]
            inter_data.input_positions[seq_idx] = inter_data.input_positions[
                seq_idx][-1:]
584
585
            inter_data.token_types[seq_idx] = inter_data.token_types[seq_idx][
                -1:]
586
587
            inter_data.query_lens[seq_idx] = 1
            inter_data.context_lens[seq_idx] = inter_data.seq_lens[seq_idx] - 1
588
589
590
591
592
593
594
595
596
597
598
599
600
601

    def _compute_for_sliding_window(self, inter_data: InterDataForSeqGroup,
                                    seq_idx: int,
                                    seq_group_metadata: SequenceGroupMetadata):
        """Update seq_len and curr_sliding_window_block for the given
        sequence data (only required by decoding) if sliding window is enabled.
        """
        curr_sliding_window_block = 0
        sliding_seq_len = inter_data.seq_lens[seq_idx]
        if not inter_data.is_prompt and self.sliding_window is not None:
            # 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.
            curr_sliding_window_block = self.sliding_window_blocks
602
603
604
605
606
607
            # number of elements in last block
            suff_len = inter_data.seq_lens[seq_idx] % self.block_size
            sliding_seq_len = min(inter_data.seq_lens[seq_idx],
                                  self.block_aligned_sliding_window + suff_len)
            if suff_len > 0:
                curr_sliding_window_block += 1
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624

        inter_data.curr_sliding_window_blocks[
            seq_idx] = curr_sliding_window_block
        inter_data.seq_lens[seq_idx] = sliding_seq_len

    def _compute_lora_input(self, inter_data: InterDataForSeqGroup,
                            seq_idx: int,
                            seq_group_metadata: SequenceGroupMetadata):
        """If LoRA is enabled, compute LoRA index and prompt mapping."""
        if not self.enable_lora:
            return

        lora_id = seq_group_metadata.lora_int_id
        if lora_id > 0:
            inter_data.lora_requests.add(seq_group_metadata.lora_request)
        query_len = inter_data.query_lens[seq_idx]
        inter_data.lora_index_mapping.append([lora_id] * query_len)
625
626
627
628
629
630
631
        sampling_params = seq_group_metadata.sampling_params
        if sampling_params and sampling_params.prompt_logprobs is not None:
            inter_data.lora_prompt_mapping.append([lora_id] * query_len)
        elif not self.chunked_prefill_enabled or seq_group_metadata.do_sample:
            inter_data.lora_prompt_mapping.append([lora_id])
        else:
            inter_data.lora_prompt_mapping.append([])
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663

    def _compute_prompt_adapter_input(
            self, inter_data: InterDataForSeqGroup,
            seq_group_metadata: SequenceGroupMetadata):
        """If prompt adapter is enabled, compute index and prompt mapping.
        """
        # Note that when is_prompt=True, we expect only one sequence
        # in the group.
        if not self.enable_prompt_adapter:
            return

        prompt_adapter_id = seq_group_metadata.prompt_adapter_id
        if prompt_adapter_id <= 0 or not inter_data.is_prompt:
            return

        # We expect only one sequence in the group when is_prompt=True.
        assert inter_data.n_seqs == 1
        query_len = inter_data.query_lens[0]
        inter_data.prompt_adapter_request = (
            seq_group_metadata.prompt_adapter_request)

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

    def _compute_multi_modal_input(self, inter_data: InterDataForSeqGroup,
                                   seq_group_metadata: SequenceGroupMetadata):
        """If multi-modal data is given, add it to the input."""
664
665
666
667
668
669
        # NOTE: mm_data only includes the subset of multi-modal items that
        # intersect with the current prefill positions.
        positions = inter_data.input_positions[0]
        mm_data, placeholder_maps = MultiModalPlaceholderMap.from_seq_group(
            seq_group_metadata,
            range(positions[0], positions[0] + len(positions)))
670
671
672
        if not mm_data:
            return

673
674
675
676
677
678
679
680
681
        if self.runner.mm_registry.has_processor(self.runner.model_config):
            mm_kwargs = mm_data
        else:
            mm_kwargs = self.multi_modal_input_mapper(
                mm_data,
                seq_group_metadata.mm_processor_kwargs,
            )

        inter_data.multi_modal_kwargs = mm_kwargs
682
        inter_data.multi_modal_placeholder_maps = placeholder_maps
683

684
        # special processing for mrope position deltas.
685
        if self.runner.model_config.uses_mrope:
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
            image_grid_thw = mm_kwargs.get("image_grid_thw", None)
            video_grid_thw = mm_kwargs.get("video_grid_thw", None)
            assert image_grid_thw is not None or video_grid_thw is not None, (
                "mrope embedding type requires multi-modal input mapper "
                "returns 'image_grid_thw' or 'video_grid_thw'.")

            hf_config = self.runner.model_config.hf_config

            inter_data.mrope_input_positions = [None] * inter_data.n_seqs
            for seq_idx in range(inter_data.n_seqs):
                seq_data = seq_group_metadata.seq_data[
                    inter_data.seq_ids[seq_idx]]
                token_ids = seq_data.get_token_ids()

                mrope_input_positions, mrope_position_delta = \
                    MRotaryEmbedding.get_input_positions(
                        token_ids,
                        image_grid_thw=image_grid_thw,
                        video_grid_thw=video_grid_thw,
                        image_token_id=hf_config.image_token_id,
                        video_token_id=hf_config.video_token_id,
                        vision_start_token_id=hf_config.vision_start_token_id,
                        vision_end_token_id=hf_config.vision_end_token_id,
                        spatial_merge_size=hf_config.vision_config.
                        spatial_merge_size,
                        context_len=inter_data.context_lens[seq_idx],
712
                        seq_len=inter_data.seq_lens[seq_idx],
713
714
715
716
717
718
                    )

                seq_data.mrope_position_delta = mrope_position_delta
                inter_data.mrope_input_positions[
                    seq_idx] = mrope_input_positions

719
    def add_seq_group(self, seq_group_metadata: SequenceGroupMetadata):
720
        """Add a sequence group to the builder."""
721
        seq_ids = seq_group_metadata.seq_data.keys()
722
723
724
725
726
727
728
        n_seqs = len(seq_ids)
        is_prompt = seq_group_metadata.is_prompt

        if is_prompt:
            assert n_seqs == 1
            self.decode_only = False

729
730
        encoder_seq_len = 0

731
        if self.runner.model_config.is_encoder_decoder:
732
733
            encoder_seq_len = seq_group_metadata.encoder_seq_data.get_len()

734
        inter_data = self.init_cached_inter_data(
735
736
737
738
            request_id=seq_group_metadata.request_id,
            seq_ids=seq_ids,
            is_prompt=is_prompt,
            block_tables=seq_group_metadata.block_tables,
739
740
            computed_block_nums=seq_group_metadata.computed_block_nums,
            reinit=True,
741
742
            reinit_use_defaults=True,
            encoder_seq_len=encoder_seq_len)
743

744
        self.inter_data_list.append(inter_data)
745

746
747
748
749
750
        for seq_idx in range(n_seqs):
            for per_seq_fn in self.per_seq_compute_fns:
                per_seq_fn(inter_data, seq_idx, seq_group_metadata)
        for per_seq_group_fn in self.per_seq_group_compute_fns:
            per_seq_group_fn(inter_data, seq_group_metadata)
751

752
753
    def _use_captured_graph(self,
                            batch_size: int,
754
                            decode_only: bool,
755
756
                            max_decode_seq_len: int,
                            max_encoder_seq_len: int = 0) -> bool:
757
        return (decode_only and not self.runner.model_config.enforce_eager
758
759
760
                and max_decode_seq_len <= self.runner.max_seq_len_to_capture
                and max_encoder_seq_len <= self.runner.max_seq_len_to_capture
                and batch_size <= self.runner.max_batchsize_to_capture)
761

762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
    def _get_cuda_graph_pad_size(self,
                                 num_seqs: int,
                                 max_decode_seq_len: int,
                                 max_encoder_seq_len: int = 0) -> int:
        """
        Determine the number of padding sequences required for running in
        CUDA graph mode. Returns -1 if CUDA graphs cannot be used.

        In the multi-step + chunked-prefill case, only the first step
        has Prefills (if any). The rest of the steps are guaranteed to be all
        decodes. In this case, we set up the padding as if all the sequences
        are decodes so we may run all steps except the first step in CUDA graph
        mode. The padding is accounted for in the multi-step `advance_step`
        family of functions.

        Args:
778
            num_seqs (int): Number of sequences scheduled to run.
779
780
781
782
783
            max_decode_seq_len (int): Greatest of all the decode sequence
                lengths. Used only in checking the viablility of using
                CUDA graphs.
            max_encoder_seq_len (int, optional): Greatest of all the encode
                sequence lengths. Defaults to 0. Used only in checking the
784
                viability of using CUDA graphs.
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
        Returns:
            int: Returns the determined number of padding sequences. If
                CUDA graphs is not viable, returns -1.
        """
        is_mscp: bool = self.runner.scheduler_config.is_multi_step and \
                    self.runner.scheduler_config.chunked_prefill_enabled
        decode_only = self.decode_only or is_mscp
        if not decode_only:
            # Early exit so we can treat num_seqs as the batch_size below.
            return -1

        # batch_size out of this function refers to the number of input
        # tokens being scheduled. This conflation of num_seqs as batch_size
        # is valid as this is a decode-only case.
        batch_size = num_seqs
        if not self._use_captured_graph(batch_size, decode_only,
                                        max_decode_seq_len,
                                        max_encoder_seq_len):
            return -1

805
        graph_batch_size = VllmConfig.get_graph_batch_size(batch_size)
806
807
808
        assert graph_batch_size >= batch_size
        return graph_batch_size - batch_size

809
    def build(self) -> ModelInputForGPU:
810
811
812
813
        """Finalize the builder intermediate data and
        create on-device tensors.
        """
        # Combine and flatten intermediate data.
814
        input_tokens = []
815
        token_types = []
816
817
818
        for inter_data in self.inter_data_list:
            for cur_input_tokens in inter_data.input_tokens:
                input_tokens.extend(cur_input_tokens)
819
820
            for cur_token_types in inter_data.token_types:
                token_types.extend(cur_token_types)
821

822
823
824
        if not input_tokens:
            # This may happen when all prefill requests hit
            # prefix caching and there is no decode request.
825
            return self.model_input_cls()
826

827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
        mrope_input_positions: Optional[List[List[int]]] = None
        if any(inter_data.mrope_input_positions is not None
               for inter_data in self.inter_data_list):
            mrope_input_positions = [[] for _ in range(3)]
            for idx in range(3):
                for inter_data in self.inter_data_list:
                    msections = inter_data.mrope_input_positions
                    if msections is None:
                        for _seq_input_positions in inter_data.input_positions:
                            mrope_input_positions[idx].extend(
                                _seq_input_positions)
                    else:
                        for _seq_mrope_input_positions in msections:
                            mrope_input_positions[idx].extend(
                                _seq_mrope_input_positions[idx])
            input_positions = None
        else:
            input_positions = []
            for inter_data in self.inter_data_list:
                for cur_input_positions in inter_data.input_positions:
                    input_positions.extend(cur_input_positions)
848

849
        seq_lens = []
850
        query_lens = []
851
        max_decode_seq_len = 0
852
        max_encoder_seq_len = 0
853
854
        for inter_data in self.inter_data_list:
            seq_lens.extend(inter_data.seq_lens)
855
            query_lens.extend(inter_data.query_lens)
856
857
858
            if not inter_data.is_prompt:
                max_decode_seq_len = max(max_decode_seq_len,
                                         max(inter_data.seq_lens))
859
                if self.runner.model_config.is_encoder_decoder:
860
861
                    max_encoder_seq_len = max(max_encoder_seq_len,
                                              inter_data.encoder_seq_len)
862

863
864
865
866
867
868
        # Mapping from request IDs to sequence IDs. Used for Jamba models
        # that manages the cache by itself.
        request_ids_to_seq_ids = {
            data.request_id: data.seq_ids
            for data in self.inter_data_list
        }
869

870
871
        cuda_graph_pad_size = self._get_cuda_graph_pad_size(
            num_seqs=len(seq_lens),
872
            max_decode_seq_len=max_decode_seq_len,
873
            max_encoder_seq_len=max_encoder_seq_len)
874

875
876
877
878
879
880
        batch_size = len(input_tokens)
        if cuda_graph_pad_size != -1:
            # If cuda graph can be used, pad tensors accordingly.
            # See `capture_model` API for more details.
            # vLLM uses cuda graph only for decoding requests.
            batch_size += cuda_graph_pad_size
881
882

        # Tokens and positions.
883
884
        if cuda_graph_pad_size:
            input_tokens.extend(itertools.repeat(0, cuda_graph_pad_size))
885
886
887
888
        assert self.runner.device is not None
        input_tokens_tensor = async_tensor_h2d(input_tokens, torch.long,
                                               self.runner.device,
                                               self.runner.pin_memory)
889
890
891
892
893
894

        token_types_tensor = async_tensor_h2d(token_types, torch.long,
                                               self.runner.device,
                                               self.runner.pin_memory) \
                                                if token_types else None

895
896
897
898
899
900
901
902
903
904
905
906
907
908
        if mrope_input_positions is not None:
            for idx in range(3):
                mrope_input_positions[idx].extend(
                    itertools.repeat(0, cuda_graph_pad_size))
            input_positions_tensor = async_tensor_h2d(mrope_input_positions,
                                                      torch.long,
                                                      self.runner.device,
                                                      self.runner.pin_memory)
        else:
            input_positions.extend(itertools.repeat(0, cuda_graph_pad_size))
            input_positions_tensor = async_tensor_h2d(input_positions,
                                                      torch.long,
                                                      self.runner.device,
                                                      self.runner.pin_memory)
909
        # Sequence and query lengths.
910
911
        if cuda_graph_pad_size:
            seq_lens.extend(itertools.repeat(1, cuda_graph_pad_size))
912
913
914

        # Attention metadata.
        attn_metadata = self.attn_metadata_builder.build(
915
            seq_lens, query_lens, cuda_graph_pad_size, batch_size)
916
917

        # LoRA data.
918
919
        lora_requests = set()
        lora_mapping = None
920
        if self.enable_lora:
921
922
923
924
925
926
            lora_requests = set(r for data in self.inter_data_list
                                for r in data.lora_requests)
            lora_index_mapping = flatten_2d_lists([
                flatten_2d_lists(inter_data.lora_index_mapping)
                for inter_data in self.inter_data_list
            ])
927
928
929
            if cuda_graph_pad_size:
                lora_index_mapping.extend(
                    itertools.repeat(0, cuda_graph_pad_size))
930
931
932
933
            lora_prompt_mapping = flatten_2d_lists([
                flatten_2d_lists(inter_data.lora_prompt_mapping)
                for inter_data in self.inter_data_list
            ])
934

935
            lora_mapping = LoRAMapping(
936
937
938
                **dict(index_mapping=lora_index_mapping,
                       prompt_mapping=lora_prompt_mapping,
                       is_prefill=not self.decode_only))
939
940

        # Prompt adapter data.
941
942
        prompt_adapter_requests: Set[PromptAdapterRequest] = set()
        prompt_adapter_mapping = None
943
        if self.enable_prompt_adapter:
944
945
946
947
948
949
950
            prompt_adapter_requests = set(
                data.prompt_adapter_request for data in self.inter_data_list
                if data.prompt_adapter_request is not None)
            prompt_adapter_index_mapping = flatten_2d_lists([
                inter_data.prompt_adapter_index_mapping
                for inter_data in self.inter_data_list
            ])
951
952
953
            if cuda_graph_pad_size:
                prompt_adapter_index_mapping.extend(
                    itertools.repeat(0, cuda_graph_pad_size))
954
955
956
957
            prompt_adapter_prompt_mapping = flatten_2d_lists([
                inter_data.prompt_adapter_prompt_mapping
                for inter_data in self.inter_data_list
            ])
958
            prompt_adapter_mapping = PromptAdapterMapping(
959
960
                prompt_adapter_index_mapping,
                prompt_adapter_prompt_mapping,
961
962
963
            )

        # Multi-modal data.
964
965
966
        multi_modal_kwargs_list = [
            data.multi_modal_kwargs for data in self.inter_data_list
            if data.multi_modal_kwargs is not None
967
        ]
968
        multi_modal_kwargs = MultiModalKwargs.batch(multi_modal_kwargs_list)
969
970
971
972

        return self.model_input_cls(
            input_tokens=input_tokens_tensor,
            input_positions=input_positions_tensor,
973
            token_types=token_types_tensor,
974
            attn_metadata=attn_metadata,
975
976
            seq_lens=seq_lens,
            query_lens=query_lens,
977
            lora_mapping=lora_mapping,
978
            lora_requests=lora_requests,
979
            multi_modal_kwargs=multi_modal_kwargs,
980
            request_ids_to_seq_ids=request_ids_to_seq_ids,
981
982
            finished_requests_ids=self.finished_requests_ids,
            prompt_adapter_mapping=prompt_adapter_mapping,
983
            prompt_adapter_requests=prompt_adapter_requests)
984
985


986
987
988
989
990
class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
    """
    Helper class for shared methods between GPU model runners.
    """
    _model_input_cls: Type[TModelInputForGPU]
991
    _builder_cls: Type[ModelInputForGPUBuilder]
992
993
994

    def __init__(
        self,
995
        vllm_config: VllmConfig,
996
        kv_cache_dtype: Optional[str] = "auto",
997
        is_driver_worker: bool = False,
998
        return_hidden_states: bool = False,
999
1000
        input_registry: InputRegistry = INPUT_REGISTRY,
        mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
1001
    ):
1002
1003
1004
1005
1006

        ModelRunnerBase.__init__(self, vllm_config)
        model_config = self.model_config
        cache_config = self.cache_config

1007
        self.is_driver_worker = is_driver_worker
1008
        self.return_hidden_states = return_hidden_states
1009

1010
        self.device = self.device_config.device
1011
        self.pin_memory = is_pin_memory_available()
1012

1013
1014
1015
1016
        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
1017
        self.max_batchsize_to_capture = VllmConfig.get_max_graph_batch_size(
1018
            self.scheduler_config.max_num_seqs)
1019
1020
1021
1022

        self.graph_runners: List[Dict[int, CUDAGraphRunner]] = [
            {} for _ in range(self.parallel_config.pipeline_parallel_size)
        ]
1023
1024
        self.graph_memory_pool: Optional[Tuple[
            int, int]] = None  # Set during graph capture.
Mor Zusman's avatar
Mor Zusman committed
1025

1026
        self.has_inner_state = model_config.has_inner_state
Mor Zusman's avatar
Mor Zusman committed
1027

1028
        # When using CUDA graph, the input block tables must be padded to
1029
        # max_seq_len_to_capture. However, creating the block table in
1030
1031
1032
        # 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
1033
        # (max batch size to capture, max seq len to capture / block size).
1034
        self.graph_block_tables = np.zeros(
1035
            (self.max_batchsize_to_capture, self.get_max_block_per_batch()),
1036
            dtype=np.int32)
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047

        # Attention-free but stateful models like Mamba need a placeholder attn
        # backend, as the attention metadata is needed to manage internal state.
        # However we must bypass attention selection altogether for some models
        # used for speculative decoding to avoid a divide-by-zero in
        # model_config.get_head_size()
        num_attn_heads = self.model_config.get_num_attention_heads(
            self.parallel_config)
        needs_attn_backend = (num_attn_heads != 0
                              or self.model_config.is_attention_free)

1048
1049
1050
1051
1052
        self.attn_backend = get_attn_backend(
            self.model_config.get_head_size(),
            self.model_config.dtype,
            self.kv_cache_dtype,
            self.block_size,
1053
            self.model_config.is_attention_free,
1054
1055
1056
1057
1058
1059
        ) if needs_attn_backend else None
        if self.attn_backend:
            self.attn_state = self.attn_backend.get_state_cls()(
                weakref.proxy(self))
        else:
            self.attn_state = CommonAttentionState(weakref.proxy(self))
1060

1061
        # Multi-modal data support
1062
1063
1064
1065
        self.input_registry = input_registry
        self.mm_registry = mm_registry
        self.multi_modal_input_mapper = mm_registry \
            .create_input_mapper(model_config)
1066
        self.mm_registry.init_mm_limits_per_prompt(self.model_config)
1067

1068
        # Lazy initialization
1069
        self.model: nn.Module  # Set after load_model
1070
1071
        # Set after load_model.
        self.lora_manager: Optional[LRUCacheWorkerLoRAManager] = None
1072
        self.prompt_adapter_manager: LRUCacheWorkerPromptAdapterManager = None
1073

1074
1075
1076
        set_cpu_offload_max_bytes(
            int(self.cache_config.cpu_offload_gb * 1024**3))

1077
1078
        # Used to cache python objects
        self.inter_data_cache: Dict[int, PyObjectCache] = {}
1079
1080
1081
1082
1083
1084
1085

        # Using the PythonizationCache in Pipeline-Parallel clobbers the
        # SequenceGroupToSample object. In Pipeline-Parallel, we have
        # more than 1 Scheduler, resulting in a potential back-to-back
        # prepare_model_inputs() call. This clobbers the cached
        # SequenceGroupToSample objects, as we reset the cache during
        # every prepare_model_inputs() call.
1086
        self.sampling_metadata_cache: SamplingMetadataCache = \
1087
1088
              SamplingMetadataCache() \
                if self.parallel_config.pipeline_parallel_size == 1 else None
1089

1090
    def load_model(self) -> None:
1091
        logger.info("Starting to load model %s...", self.model_config.model)
1092
        with DeviceMemoryProfiler() as m:
1093
            self.model = get_model(vllm_config=self.vllm_config)
1094
1095

        self.model_memory_usage = m.consumed_memory
1096
1097
        logger.info("Loading model weights took %.4f GB",
                    self.model_memory_usage / float(2**30))
1098
1099

        if self.lora_config:
1100
            assert supports_lora(
1101
                self.model
1102
            ), f"{self.model.__class__.__name__} does not support LoRA yet."
1103

1104
1105
1106
            if supports_multimodal(self.model):
                logger.warning("Regarding multimodal models, vLLM currently "
                               "only supports adding LoRA to language model.")
1107
1108
1109
1110
1111
1112
1113
            # It's necessary to distinguish between the max_position_embeddings
            # of VLMs and LLMs.
            if hasattr(self.model.config, "max_position_embeddings"):
                max_pos_embeddings = self.model.config.max_position_embeddings
            else:
                max_pos_embeddings = (
                    self.model.config.text_config.max_position_embeddings)
1114

1115
1116
            self.lora_manager = LRUCacheWorkerLoRAManager(
                self.scheduler_config.max_num_seqs,
1117
1118
1119
1120
1121
1122
                self.scheduler_config.max_num_batched_tokens,
                self.vocab_size,
                self.lora_config,
                self.device,
                self.model.embedding_modules,
                self.model.embedding_padding_modules,
1123
                max_position_embeddings=max_pos_embeddings,
1124
            )
1125
            self.model = self.lora_manager.create_lora_manager(self.model)
1126

1127
1128
1129
1130
1131
1132
1133
1134
1135
        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))

1136
        if self.kv_cache_dtype == "fp8" and current_platform.is_rocm():
1137
1138
1139
            # Currently only ROCm accepts kv-cache scaling factors
            # via quantization_param_path and this will be deprecated
            # in the future.
1140
1141
            if self.model_config.quantization_param_path is not None:
                if callable(getattr(self.model, "load_kv_cache_scales", None)):
1142
1143
1144
1145
1146
1147
                    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)
1148
1149
                    self.model.load_kv_cache_scales(
                        self.model_config.quantization_param_path)
1150
1151
                    logger.info("Loaded KV cache scaling factors from %s",
                                self.model_config.quantization_param_path)
1152
                else:
1153
1154
1155
1156
                    raise RuntimeError(
                        "Using FP8 KV cache and scaling factors provided but "
                        "model %s does not support loading scaling factors.",
                        self.model.__class__)
1157
            else:
1158
1159
1160
1161
                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!")
1162

1163
1164
        if self.vllm_config.compilation_config.level ==\
            CompilationLevel.DYNAMO_AS_IS and supports_dynamo():
1165
1166
            backend = self.vllm_config.compilation_config.init_backend(
                self.vllm_config)
1167
1168
1169
            self.model = torch.compile(
                self.model,
                fullgraph=envs.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE,
1170
                backend=backend)
1171

1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
    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,
        )

1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
    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,
        )

1196
1197
    def get_max_block_per_batch(self) -> int:
        block_size = self.block_size
1198
        return (self.max_seq_len_to_capture + block_size - 1) // block_size
1199

1200
    def _prepare_model_input_tensors(
1201
1202
        self,
        seq_group_metadata_list: List[SequenceGroupMetadata],
Mor Zusman's avatar
Mor Zusman committed
1203
        finished_requests_ids: Optional[List[str]] = None
1204
1205
1206
1207
    ) -> 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.
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218

        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.
        """
1219
        builder = self._builder_cls(weakref.proxy(self), finished_requests_ids)
1220
        for seq_group_metadata in seq_group_metadata_list:
1221
            builder.add_seq_group(seq_group_metadata)
1222
1223
1224

        builder.reset_cached_inter_data()

1225
        return builder.build()  # type: ignore
1226

1227
1228
1229
    @torch.inference_mode()
    def profile_run(self) -> None:
        # Enable top-k sampling to reflect the accurate memory usage.
1230
        sampling_params = SamplingParams(top_p=0.99, top_k=self.vocab_size - 1)
1231
1232
        max_num_batched_tokens = self.scheduler_config.max_num_batched_tokens
        max_num_seqs = self.scheduler_config.max_num_seqs
1233
1234
1235
1236
        # 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.
1237
1238
        dummy_lora_requests: List[LoRARequest] = []
        dummy_lora_requests_per_seq: List[LoRARequest] = []
1239
        if self.lora_config:
1240
            assert self.lora_manager is not None
1241
1242
1243
1244
1245
1246
            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,
1247
                        lora_path="/not/a/real/path",
1248
1249
1250
1251
1252
1253
1254
1255
                    )
                    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)
                ]
1256

1257
1258
1259
        # Profile memory usage with max_num_sequences sequences and the total
        # number of tokens equal to max_num_batched_tokens.
        seqs: List[SequenceGroupMetadata] = []
1260
1261
        # Additional GPU memory may be needed for multi-modal encoding, which
        # needs to be accounted for when calculating the GPU blocks for
1262
1263
1264
1265
        # 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.
1266

1267
1268
        max_mm_tokens = self.mm_registry.get_max_multimodal_tokens(
            self.model_config)
1269
        if max_mm_tokens > 0:
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
            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

1281
        batch_size = 0
1282
1283
1284
        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))
1285
            batch_size += seq_len
1286

1287
            dummy_data = self.input_registry \
1288
1289
1290
                .dummy_data_for_profiling(self.model_config,
                                          seq_len,
                                          self.mm_registry)
1291

1292
1293
1294
            seq = SequenceGroupMetadata(
                request_id=str(group_id),
                is_prompt=True,
1295
                seq_data={group_id: dummy_data.seq_data},
1296
1297
                sampling_params=sampling_params,
                block_tables=None,
1298
1299
                lora_request=dummy_lora_requests_per_seq[group_id]
                if dummy_lora_requests_per_seq else None,
1300
1301
                multi_modal_data=dummy_data.multi_modal_data,
                multi_modal_placeholders=dummy_data.multi_modal_placeholders,
1302
1303
1304
1305
1306
            )
            seqs.append(seq)

        # Run the model with the dummy inputs.
        num_layers = self.model_config.get_num_layers(self.parallel_config)
1307
1308
1309
1310
        # use an empty tensor instead of `None`` to force Dynamo to pass
        # it by reference, rather by specializing on the value ``None``.
        # the `dtype` argument does not matter, and we use `float32` as
        # a placeholder (it has wide hardware support).
1311
1312
1313
        # it is important to create tensors inside the loop, rather than
        # multiplying the list, to avoid Dynamo from treating them as
        # tensor aliasing.
1314
1315
        kv_caches = [
            torch.tensor([], dtype=torch.float32, device=self.device)
1316
1317
            for _ in range(num_layers)
        ]
Mor Zusman's avatar
Mor Zusman committed
1318
1319
1320
        finished_requests_ids = [seq.request_id for seq in seqs]
        model_input = self.prepare_model_input(
            seqs, finished_requests_ids=finished_requests_ids)
1321
1322
1323
1324
1325
1326
        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)
1327

1328
        self.execute_model(model_input, kv_caches, intermediate_tensors)
1329
        torch.cuda.synchronize()
1330
1331
        return

1332
    def remove_all_loras(self):
1333
1334
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
1335
        self.lora_manager.remove_all_adapters()
1336

1337
    def set_active_loras(self, lora_requests: Set[LoRARequest],
1338
1339
1340
                         lora_mapping: LoRAMapping) -> None:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
1341
        self.lora_manager.set_active_adapters(lora_requests, lora_mapping)
1342
1343
1344
1345

    def add_lora(self, lora_request: LoRARequest) -> bool:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
1346
        return self.lora_manager.add_adapter(lora_request)
1347
1348
1349
1350

    def remove_lora(self, lora_id: int) -> bool:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
1351
        return self.lora_manager.remove_adapter(lora_id)
1352
1353
1354
1355

    def pin_lora(self, lora_id: int) -> bool:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
1356
        return self.lora_manager.pin_adapter(lora_id)
1357
1358
1359
1360

    def list_loras(self) -> Set[int]:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
        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()
1396

1397
    @torch.inference_mode()
1398
    def capture_model(self, kv_caches: List[List[torch.Tensor]]) -> None:
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
        """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.
        """
1411
        assert not self.model_config.enforce_eager
1412
        logger.info("Capturing cudagraphs for decoding. This may lead to "
1413
1414
1415
                    "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.")
1416
1417
1418
1419
        logger.info("If out-of-memory error occurs during cudagraph capture,"
                    " consider decreasing `gpu_memory_utilization` or "
                    "switching to eager mode. You can also reduce the "
                    "`max_num_seqs` as needed to decrease memory usage.")
1420
        start_time = time.perf_counter()
1421
        start_free_gpu_memory = torch.cuda.mem_get_info()[0]
1422
1423

        # Prepare dummy inputs. These will be reused for all batch sizes.
1424
        max_batch_size = self.max_batchsize_to_capture
1425
1426
        input_tokens = torch.zeros(max_batch_size, dtype=torch.long).cuda()
        input_positions = torch.zeros(max_batch_size, dtype=torch.long).cuda()
1427
        if self.model_config.uses_mrope:
1428
            input_positions = torch.tile(input_positions, (3, 1))
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
        # Prepare dummy previous_hidden_states only if needed by the model.
        # This is used by draft models such as EAGLE.
        previous_hidden_states = None
        if "previous_hidden_states" in inspect.signature(
                self.model.forward).parameters:
            previous_hidden_states = torch.empty(
                [max_batch_size,
                 self.model_config.get_hidden_size()],
                dtype=self.model_config.dtype,
                device=self.device)

1440
1441
1442
1443
1444
1445
        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)
1446

1447
1448
        with self.attn_state.graph_capture(
                max_batch_size), graph_capture() as graph_capture_context:
1449
1450
            # NOTE: Capturing the largest batch size first may help reduce the
            # memory usage of CUDA graph.
1451
1452
            for virtual_engine in range(
                    self.parallel_config.pipeline_parallel_size):
1453
1454
                for batch_size in \
                    self.vllm_config.compilation_config.capture_sizes:
1455
1456
                    attn_metadata = (
                        self.attn_state.graph_capture_get_metadata_for_batch(
1457
1458
                            batch_size,
                            is_encoder_decoder_model=self.model_config.
1459
                            is_encoder_decoder))
1460
1461
1462

                    if self.lora_config:
                        lora_mapping = LoRAMapping(
1463
1464
1465
                            **dict(index_mapping=[0] * batch_size,
                                   prompt_mapping=[0] * batch_size,
                                   is_prefill=False))
1466
1467
                        self.set_active_loras(set(), lora_mapping)

1468
1469
1470
1471
1472
1473
1474
                    if self.prompt_adapter_config:
                        prompt_adapter_mapping = PromptAdapterMapping(
                            [-1] * batch_size,
                            [-1] * batch_size,
                        )
                        self.set_active_prompt_adapters(
                            set(), prompt_adapter_mapping)
1475
                    graph_runner = CUDAGraphRunner(
1476
                        self.model, self.attn_backend.get_name(),
1477
                        self.attn_state.graph_clone(batch_size),
1478
                        self.model_config.is_encoder_decoder)
1479

Mor Zusman's avatar
Mor Zusman committed
1480
1481
                    capture_inputs = {
                        "input_ids":
1482
                        input_tokens[:batch_size],
Mor Zusman's avatar
Mor Zusman committed
1483
                        "positions":
1484
                        input_positions[..., :batch_size],
Mor Zusman's avatar
Mor Zusman committed
1485
                        "intermediate_inputs":
1486
1487
                        intermediate_inputs[:batch_size]
                        if intermediate_inputs is not None else None,
Mor Zusman's avatar
Mor Zusman committed
1488
                        "kv_caches":
1489
                        kv_caches[virtual_engine],
Mor Zusman's avatar
Mor Zusman committed
1490
                        "attn_metadata":
1491
                        attn_metadata,
Mor Zusman's avatar
Mor Zusman committed
1492
1493
1494
1495
1496
                        "memory_pool":
                        self.graph_memory_pool,
                        "stream":
                        graph_capture_context.stream
                    }
1497
1498
1499
1500
1501
                    if previous_hidden_states is not None:
                        capture_inputs[
                            "previous_hidden_states"] = previous_hidden_states[:
                                                                               batch_size]

1502
                    if self.has_inner_state:
Mor Zusman's avatar
Mor Zusman committed
1503
1504
1505
1506
1507
1508
                        # 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)
                        })
1509
                    if self.model_config.is_encoder_decoder:
1510
1511
1512
1513
1514
                        # add the additional inputs to capture for
                        # encoder-decoder models.
                        self._update_inputs_to_capture_for_enc_dec_model(
                            capture_inputs)

1515
                    with set_forward_context(attn_metadata, self.vllm_config):
1516
                        graph_runner.capture(**capture_inputs)
1517
1518
1519
                    self.graph_memory_pool = graph_runner.graph.pool()
                    self.graph_runners[virtual_engine][batch_size] = (
                        graph_runner)
1520
1521

        end_time = time.perf_counter()
1522
        end_free_gpu_memory = torch.cuda.mem_get_info()[0]
1523
        elapsed_time = end_time - start_time
1524
        cuda_graph_size = start_free_gpu_memory - end_free_gpu_memory
1525
        # This usually takes < 10 seconds.
1526
1527
        logger.info("Graph capturing finished in %.0f secs, took %.2f GiB",
                    elapsed_time, cuda_graph_size / GiB_bytes)
1528

1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
    def _update_inputs_to_capture_for_enc_dec_model(self,
                                                    capture_inputs: Dict[str,
                                                                         Any]):
        """
        Updates the set of input tensors needed for CUDA graph capture in an
        encoder-decoder model.

        This method modifies the provided `capture_inputs` dictionary by
        adding tensors specific to encoder-decoder specific models that
        need to be captured for CUDA Graph replay.
        """
        # During the decode phase encoder_input_ids and encoder_positions are
        # unset. Do the same thing for graph capture.
        capture_inputs["encoder_input_ids"] = torch.tensor(
            [], dtype=torch.long).cuda()
        capture_inputs["encoder_positions"] = torch.tensor(
            [], dtype=torch.long).cuda()

1547
1548
1549
1550
    @property
    def vocab_size(self) -> int:
        return self.model_config.get_vocab_size()

1551

1552
1553
1554
1555
1556
1557
class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
    """
    GPU model runner with sampling step.
    """
    _model_input_cls: Type[ModelInputForGPUWithSamplingMetadata] = (
        ModelInputForGPUWithSamplingMetadata)
1558
    _builder_cls: Type[ModelInputForGPUBuilder] = ModelInputForGPUBuilder
1559
1560
1561
1562
1563

    def make_model_input_from_broadcasted_tensor_dict(
        self,
        tensor_dict: Dict[str, Any],
    ) -> ModelInputForGPUWithSamplingMetadata:
1564
        model_input = \
1565
1566
1567
            ModelInputForGPUWithSamplingMetadata.from_broadcasted_tensor_dict(
                tensor_dict,
                attn_backend=self.attn_backend,
1568
1569
            )
        return model_input
1570
1571
1572
1573

    def prepare_model_input(
        self,
        seq_group_metadata_list: List[SequenceGroupMetadata],
1574
        virtual_engine: int = 0,
1575
        finished_requests_ids: Optional[List[str]] = None,
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
    ) -> 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
1591
            seq_group_metadata_list, finished_requests_ids)
1592
1593
1594
1595
1596
1597
        if get_pp_group().is_last_rank:
            # Sampling metadata is only required for the final pp group
            generators = self.get_generators(finished_requests_ids)
            sampling_metadata = SamplingMetadata.prepare(
                seq_group_metadata_list, model_input.seq_lens,
                model_input.query_lens, self.device, self.pin_memory,
1598
                generators, self.sampling_metadata_cache)
1599
1600
        else:
            sampling_metadata = None
1601
1602
1603
1604
        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,
1605
1606
                                   is_prompt=is_prompt,
                                   virtual_engine=virtual_engine)
1607
1608

    @torch.inference_mode()
1609
    @dump_input_when_exception(exclude_args=[0], exclude_kwargs=["self"])
1610
1611
1612
1613
    def execute_model(
        self,
        model_input: ModelInputForGPUWithSamplingMetadata,
        kv_caches: List[torch.Tensor],
1614
        intermediate_tensors: Optional[IntermediateTensors] = None,
1615
        num_steps: int = 1,
1616
    ) -> Optional[Union[List[SamplerOutput], IntermediateTensors]]:
1617
1618
1619
        if num_steps > 1:
            raise ValueError("num_steps > 1 is not supported in ModelRunner")

1620
1621
1622
1623
1624
1625
        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)

1626
1627
1628
1629
1630
1631
1632
        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)

1633
        self.attn_state.begin_forward(model_input)
1634

1635
1636
1637
1638
        # 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
1639
1640
1641
        # TODO(andoorve): We can remove this once all
        # virtual engines share the same kv cache.
        virtual_engine = model_input.virtual_engine
1642
1643
1644
        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]
1645
1646
            model_executable = self.graph_runners[virtual_engine][
                graph_batch_size]
1647
1648
1649
        else:
            model_executable = self.model

1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
        # Receive KV cache in distributed KV cache transfer setting
        # In disagg prefill setting, it will also recv hidden states and bypass
        # model forwarding
        # In KV cache database setting, it will change the model input so that
        # we can skip prefilling on tokens that successfully received KV caches
        # NOTE: The receive operation is blocking
        bypass_model_exec = False
        if self.need_recv_kv(model_input, kv_caches):
            hidden_or_intermediate_states, bypass_model_exec, model_input = \
                get_kv_transfer_group().recv_kv_caches_and_hidden_states(
                    # model is used to know which layer the current worker
                    # is working on, so that we can receive KV for only those
                    # layers.
                    model_executable,
                    model_input,
                    kv_caches=kv_caches
                )

1668
        multi_modal_kwargs = model_input.multi_modal_kwargs or {}
Mor Zusman's avatar
Mor Zusman committed
1669
1670
1671
        seqlen_agnostic_kwargs = {
            "finished_requests_ids": model_input.finished_requests_ids,
            "request_ids_to_seq_ids": model_input.request_ids_to_seq_ids,
1672
        } if self.has_inner_state else {}
1673
1674
1675
1676
1677
1678
        if (self.observability_config is not None
                and self.observability_config.collect_model_forward_time):
            model_forward_start = torch.cuda.Event(enable_timing=True)
            model_forward_end = torch.cuda.Event(enable_timing=True)
            model_forward_start.record()

1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
        if not bypass_model_exec:
            with set_forward_context(model_input.attn_metadata,
                                     self.vllm_config):
                hidden_or_intermediate_states = model_executable(
                    input_ids=model_input.input_tokens,
                    positions=model_input.input_positions,
                    kv_caches=kv_caches,
                    attn_metadata=model_input.attn_metadata,
                    intermediate_tensors=intermediate_tensors,
                    **MultiModalKwargs.as_kwargs(multi_modal_kwargs,
                                                 device=self.device),
                    **seqlen_agnostic_kwargs)
1691

1692
1693
1694
1695
        if (self.observability_config is not None
                and self.observability_config.collect_model_forward_time):
            model_forward_end.record()

1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
        # Sending KV cache in distributed KV cache transfer setting
        # NOTE: the send operation is non-blocking
        if self.need_send_kv(model_input, kv_caches):
            get_kv_transfer_group().send_kv_caches_and_hidden_states(
                # model_executable is used to know which layer the current
                # worker is working on, so that we can send KV for only those
                # layers.
                model_executable,
                model_input,
                kv_caches,
                hidden_or_intermediate_states,
            )

1709
1710
        # Compute the logits in the last pipeline stage.
        if not get_pp_group().is_last_rank:
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
            if (self.is_driver_worker
                    and hidden_or_intermediate_states is not None
                    and isinstance(hidden_or_intermediate_states,
                                   IntermediateTensors)
                    and self.observability_config is not None
                    and self.observability_config.collect_model_forward_time):
                model_forward_end.synchronize()
                model_forward_time = model_forward_start.elapsed_time(
                    model_forward_end)
                orig_model_forward_time = 0.0
                if intermediate_tensors is not None:
                    orig_model_forward_time = intermediate_tensors.tensors.get(
                        "model_forward_time", torch.tensor(0.0)).item()
                hidden_or_intermediate_states.tensors["model_forward_time"] = (
                    torch.tensor(model_forward_time + orig_model_forward_time))
1726
1727
1728
            return hidden_or_intermediate_states

        logits = self.model.compute_logits(hidden_or_intermediate_states,
1729
1730
1731
                                           model_input.sampling_metadata)

        if not self.is_driver_worker:
1732
            return []
1733

1734
1735
        if model_input.async_callback is not None:
            model_input.async_callback()
1736

1737
1738
1739
1740
1741
        # Sample the next token.
        output: SamplerOutput = self.model.sample(
            logits=logits,
            sampling_metadata=model_input.sampling_metadata,
        )
1742
1743
1744
1745
1746
1747
        if (self.observability_config is not None
                and self.observability_config.collect_model_forward_time
                and output is not None):
            model_forward_end.synchronize()
            model_forward_time = model_forward_start.elapsed_time(
                model_forward_end)
1748
1749
1750
1751
            orig_model_forward_time = 0.0
            if intermediate_tensors is not None:
                orig_model_forward_time = intermediate_tensors.tensors.get(
                    "model_forward_time", torch.tensor(0.0)).item()
1752
1753
1754
1755
            # If there are multiple workers, we are still tracking the latency
            # from the start time of the driver worker to the end time of the
            # driver worker. The model forward time will then end up covering
            # the communication time as well.
1756
1757
            output.model_forward_time = (orig_model_forward_time +
                                         model_forward_time)
1758
1759
1760

        if self.return_hidden_states:
            # we only need to pass hidden states of most recent token
1761
1762
            assert model_input.sampling_metadata is not None
            indices = model_input.sampling_metadata.selected_token_indices
1763
            if model_input.is_prompt:
1764
1765
                hidden_states = hidden_or_intermediate_states.index_select(
                    0, indices)
1766
                output.prefill_hidden_states = hidden_or_intermediate_states
1767
            elif decode_meta.use_cuda_graph:
1768
1769
1770
                hidden_states = hidden_or_intermediate_states[:len(indices)]
            else:
                hidden_states = hidden_or_intermediate_states
1771

1772
1773
            output.hidden_states = hidden_states

1774
        return [output]
1775

1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
    def need_recv_kv(self, model_input, kv_caches) -> bool:
        """Check if we need to receive kv-cache from the other worker.
        We need to receive KV when
            1. current vLLM instance is KV cache consumer/decode vLLM instance
            2. this batch is not a profiling run
            3. this batch is a prefill run
            
        Args:
            model_input: input to the model executable
            kv_caches: vLLM's paged memory
        """

youkaichao's avatar
youkaichao committed
1788
1789
1790
        if self.vllm_config.kv_transfer_config is None:
            return False

1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
        prefill_meta = model_input.attn_metadata.prefill_metadata

        # check if the current run is profiling
        is_profile_run = (kv_caches[0].numel() == 0)
        # check if the current run is prefill
        is_prefill_run = prefill_meta is not None

        return self.vllm_config.kv_transfer_config.is_kv_consumer and (
            not is_profile_run) and is_prefill_run

    def need_send_kv(self, model_input, kv_caches) -> bool:
        """Check if we need to send kv-cache to the other worker.
        We need to send KV when
            1. current vLLM instance is KV cache producer/prefill vLLM instance
            2. this batch is not a profiling run
            3. this batch is a prefill run
            
        Args:
            model_input: input to the model executable
            kv_caches: vLLM's paged memory
        """

youkaichao's avatar
youkaichao committed
1813
1814
1815
        if self.vllm_config.kv_transfer_config is None:
            return False

1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
        prefill_meta = model_input.attn_metadata.prefill_metadata

        # check if the current run is profiling
        is_profile_run = (kv_caches[0].numel() == 0)
        # check if the current run is prefill
        is_prefill_run = prefill_meta is not None

        return self.vllm_config.kv_transfer_config.is_kv_producer and (
            not is_profile_run) and is_prefill_run

1826

1827
1828
1829
# NOTE: this is nn.Module so the profiler can properly capture/group
#  kernels calls made within the graph
class CUDAGraphRunner(nn.Module):
1830

1831
    def __init__(self, model: nn.Module, backend_name: str,
1832
                 attn_state: AttentionState, is_encoder_decoder_model: bool):
1833
        super().__init__()
1834
        self.model = model
1835
        self.backend_name = backend_name
1836
        self.attn_state = attn_state
1837

1838
1839
1840
        self.input_buffers: Dict[str, torch.Tensor] = {}
        self.output_buffers: Dict[str, torch.Tensor] = {}

1841
        self._graph: Optional[torch.cuda.CUDAGraph] = None
1842
        self._is_encoder_decoder_model = is_encoder_decoder_model
1843
1844
1845
1846
1847
1848

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

1849
1850
1851
1852
    def capture(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1853
        intermediate_inputs: Optional[IntermediateTensors],
1854
1855
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
1856
1857
        memory_pool: Optional[Tuple[int, int]],
        stream: torch.cuda.Stream,
1858
        **kwargs,
1859
    ):
1860
        assert self._graph is None
1861
        # Run the model a few times without capturing the graph.
1862
1863
        # This is to make sure that the captured graph does not include the
        # kernel launches for initial benchmarking (e.g., Triton autotune).
1864
        # Note one iteration is not enough for torch.compile
1865
1866
        for _ in range(_NUM_WARMUP_ITERS):
            self.model(
1867
1868
1869
1870
1871
                input_ids=input_ids,
                positions=positions,
                kv_caches=kv_caches,
                attn_metadata=attn_metadata,
                intermediate_tensors=intermediate_inputs,
1872
1873
                **kwargs,
            )
1874
1875
        # Wait for the warm up operations to finish before proceeding with
        # Graph Capture.
1876
1877
1878
1879
        torch.cuda.synchronize()
        # Capture the graph.
        self._graph = torch.cuda.CUDAGraph()
        with torch.cuda.graph(self._graph, pool=memory_pool, stream=stream):
1880
            output_hidden_or_intermediate_states = self.model(
1881
1882
1883
1884
1885
                input_ids=input_ids,
                positions=positions,
                kv_caches=kv_caches,
                attn_metadata=attn_metadata,
                intermediate_tensors=intermediate_inputs,
1886
                **kwargs,
1887
            )
1888
1889
1890

            if isinstance(output_hidden_or_intermediate_states, torch.Tensor):
                hidden_or_intermediate_states = weak_ref_tensor(
1891
                    output_hidden_or_intermediate_states)
1892
1893
1894
1895
1896
1897
1898
1899
            elif isinstance(output_hidden_or_intermediate_states,
                            IntermediateTensors):
                hidden_or_intermediate_states = IntermediateTensors(
                    tensors={
                        key: weak_ref_tensor(value)
                        for key, value in
                        output_hidden_or_intermediate_states.tensors.items()
                    })
1900
1901

            del output_hidden_or_intermediate_states
1902
            # make sure `output_hidden_or_intermediate_states` is deleted
1903
1904
            # in the graph's memory pool
            gc.collect()
1905
1906
1907
        torch.cuda.synchronize()

        # Save the input and output buffers.
1908
        self.input_buffers = {
1909
1910
1911
1912
1913
1914
1915
1916
            "input_ids":
            input_ids,
            "positions":
            positions,
            "kv_caches":
            kv_caches,
            **self.attn_state.get_graph_input_buffers(
                attn_metadata, self._is_encoder_decoder_model),
1917
1918
            **kwargs,
        }
1919
1920
1921
1922
1923
1924
1925
1926
        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
1927
1928
1929
1930
1931

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1932
1933
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
1934
        intermediate_tensors: Optional[IntermediateTensors],
1935
        **kwargs,
1936
1937
1938
1939
1940
    ) -> 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.
1941
1942
        self.input_buffers["input_ids"].copy_(input_ids, non_blocking=True)
        self.input_buffers["positions"].copy_(positions, non_blocking=True)
1943

1944
        if self.backend_name != "NO_ATTENTION":
1945
1946
1947
            self.input_buffers["slot_mapping"].copy_(
                attn_metadata.slot_mapping, non_blocking=True)

1948
1949
        self.attn_state.prepare_graph_input_buffers(
            self.input_buffers, attn_metadata, self._is_encoder_decoder_model)
1950

Mor Zusman's avatar
Mor Zusman committed
1951
1952
1953
        if "seqlen_agnostic_capture_inputs" in self.input_buffers:
            self.model.copy_inputs_before_cuda_graphs(self.input_buffers,
                                                      **kwargs)
1954
1955
1956
1957
1958

        if "previous_hidden_states" in self.input_buffers:
            self.input_buffers["previous_hidden_states"].copy_(
                kwargs["previous_hidden_states"], non_blocking=True)

1959
1960
        if intermediate_tensors is not None:
            for key in intermediate_tensors.tensors:
1961
                if key != "model_execute_time" and key != "model_forward_time":
1962
1963
                    self.input_buffers[key].copy_(intermediate_tensors[key],
                                                  non_blocking=True)
1964
1965
1966
1967
1968
1969
        if self._is_encoder_decoder_model:
            self.input_buffers["encoder_input_ids"].copy_(
                kwargs['encoder_input_ids'], non_blocking=True)
            self.input_buffers["encoder_positions"].copy_(
                kwargs['encoder_positions'], non_blocking=True)

1970
1971
1972
        # Run the graph.
        self.graph.replay()
        # Return the output tensor.
1973
1974
1975
1976
        if get_pp_group().is_last_rank:
            return self.output_buffers["hidden_states"]

        return self.output_buffers