model_runner.py 89 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 contextlib import contextmanager
9
from dataclasses import dataclass
10
11
from typing import (TYPE_CHECKING, Any, Callable, Dict, List, Optional, Set,
                    Tuple, Type, TypeVar, Union)
12

13
import numpy as np
14
import torch
15
import torch.distributed
16
import torch.nn as nn
17
from tqdm import tqdm
18

19
import vllm.envs as envs
20
from vllm.attention import AttentionMetadata, get_attn_backend
21
from vllm.attention.backends.abstract import AttentionState
22
from vllm.attention.backends.utils import CommonAttentionState
23
from vllm.config import CompilationLevel, VllmConfig
24
from vllm.core.scheduler import SchedulerOutputs
25
from vllm.distributed import get_kv_transfer_group, get_pp_group
26
27
from vllm.distributed.parallel_state import (get_tensor_model_parallel_rank,
                                             graph_capture)
28
from vllm.forward_context import set_forward_context
29
from vllm.inputs import INPUT_REGISTRY, InputRegistry
30
from vllm.logger import init_logger
31
32
33
from vllm.lora.layers import LoRAMapping
from vllm.lora.request import LoRARequest
from vllm.lora.worker_manager import LRUCacheWorkerLoRAManager
34
from vllm.model_executor import SamplingMetadata, SamplingMetadataCache
35
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
36
from vllm.model_executor.layers.sampler import SamplerOutput
37
from vllm.model_executor.model_loader import get_model
38
from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
39
from vllm.model_executor.models import supports_lora, supports_multimodal
40
from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
41
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
42
                             MultiModalKwargs, MultiModalPlaceholderMap,
43
                             MultiModalRegistry)
44
from vllm.platforms import current_platform
45
46
47
48
from vllm.prompt_adapter.layers import PromptAdapterMapping
from vllm.prompt_adapter.request import PromptAdapterRequest
from vllm.prompt_adapter.worker_manager import (
    LRUCacheWorkerPromptAdapterManager)
49
from vllm.sampling_params import SamplingParams
50
from vllm.sequence import IntermediateTensors, SequenceGroupMetadata
51
52
53
54
from vllm.utils import (DeviceMemoryProfiler, GiB_bytes, PyObjectCache,
                        async_tensor_h2d, flatten_2d_lists,
                        is_pin_memory_available, supports_dynamo,
                        weak_ref_tensor)
55
from vllm.worker.model_runner_base import (
56
    ModelRunnerBase, ModelRunnerInputBase, ModelRunnerInputBuilderBase,
57
58
59
    _add_attn_metadata_broadcastable_dict,
    _add_sampling_metadata_broadcastable_dict,
    _init_attn_metadata_from_tensor_dict,
60
    _init_sampling_metadata_from_tensor_dict, dump_input_when_exception)
61
62
63

if TYPE_CHECKING:
    from vllm.attention.backends.abstract import AttentionBackend
64
65
66

logger = init_logger(__name__)

67
LORA_WARMUP_RANK = 8
68

69
_NUM_WARMUP_ITERS = 2
70

71
72
TModelInputForGPU = TypeVar('TModelInputForGPU', bound="ModelInputForGPU")

73
74
75
76
# 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

77

78
@dataclass(frozen=True)
79
80
81
82
83
84
85
86
87
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
88
    token_types: Optional[torch.Tensor] = None
89
90
91
92
93
    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
94
95
    prompt_adapter_mapping: Optional[PromptAdapterMapping] = None
    prompt_adapter_requests: Optional[Set[PromptAdapterRequest]] = None
96
    multi_modal_kwargs: Optional[BatchedTensorInputs] = None
Mor Zusman's avatar
Mor Zusman committed
97
98
    request_ids_to_seq_ids: Optional[Dict[str, List[int]]] = None
    finished_requests_ids: Optional[List[str]] = None
99
    virtual_engine: int = 0
100
    async_callback: Optional[Callable] = None
101
102
    seq_group_metadata_list: Optional[List[SequenceGroupMetadata]] = None
    scheduler_outputs: Optional[SchedulerOutputs] = None
103
104
105
106
107
108
109
110

    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,
111
112
            "prompt_adapter_mapping": self.prompt_adapter_mapping,
            "prompt_adapter_requests": self.prompt_adapter_requests,
113
            "virtual_engine": self.virtual_engine,
Mor Zusman's avatar
Mor Zusman committed
114
115
            "request_ids_to_seq_ids": self.request_ids_to_seq_ids,
            "finished_requests_ids": self.finished_requests_ids,
116
117
118
        }
        _add_attn_metadata_broadcastable_dict(tensor_dict, self.attn_metadata)
        return tensor_dict
119
120

    @classmethod
121
122
123
124
125
126
127
128
129
130
    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)

131
132
133
134
135
136
137
138
139
140
141
142
    # 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})

143

144
@dataclass(frozen=True)
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
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,
161
162
            "prompt_adapter_mapping": self.prompt_adapter_mapping,
            "prompt_adapter_requests": self.prompt_adapter_requests,
163
            "virtual_engine": self.virtual_engine,
Mor Zusman's avatar
Mor Zusman committed
164
165
            "request_ids_to_seq_ids": self.request_ids_to_seq_ids,
            "finished_requests_ids": self.finished_requests_ids,
166
167
168
169
170
        }
        _add_attn_metadata_broadcastable_dict(tensor_dict, self.attn_metadata)
        _add_sampling_metadata_broadcastable_dict(tensor_dict,
                                                  self.sampling_metadata)
        return tensor_dict
171

172
173
174
175
176
177
178
179
180
181
182
183
184
    @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)


185
class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
186
187
    """Build ModelInputForGPU from SequenceGroupMetadata."""

188
189
190
    # 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.
191
192
    class InterDataForSeqGroup:
        """Intermediate data for the current sequence group."""
193

194
195
196
        def simple_reinit(self):
            self.input_tokens[0].clear()  # type: ignore
            self.input_positions[0].clear()  # type: ignore
197
            self.token_types[0].clear()  # type: ignore
198
            self.mrope_input_positions = None  # type: ignore
199
200
201
202
203
204
205
206
207
208
209
            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

210
211
212
213
214
215
216
217
218
219
220
221
222
223
        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,
224
            token_types: Optional[List[List[int]]] = None,
225
            mrope_input_positions: Optional[List[List[List[int]]]] = None,
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249

            # 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.
250
            multi_modal_kwargs: Optional[MultiModalKwargs] = None,
251
252
            multi_modal_placeholder_maps: Optional[Dict[
                str, MultiModalPlaceholderMap]] = None,
253
254
255

            # Whether the prefix cache is hit (prefill only).
            prefix_cache_hit: bool = False,
256
257
            reinit: bool = False,
            reinit_use_defaults: bool = False,
258
            encoder_seq_len: int = 0,
259
        ):
260
261
262
263
264
265
266
            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

267
268
269
270
271
            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
272
            self.encoder_seq_len = encoder_seq_len
273

274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
            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()

290
291
292
293
294
295
                    if token_types:
                        self.token_types = token_types
                    else:
                        for seq_id in range(len(self.seq_ids)):
                            self.token_types[seq_id].clear()

296
297
                    self.mrope_input_positions = None

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
356
357
358
                    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 []
359
                self.token_types = token_types or []
360
                self.mrope_input_positions = mrope_input_positions or None
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
                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
378
            self.multi_modal_kwargs = multi_modal_kwargs
379
            self.multi_modal_placeholder_maps = multi_modal_placeholder_maps
380
381
            self.prefix_cache_hit = prefix_cache_hit

382
383
            self.n_seqs = len(self.seq_ids)

384
385
            if not reinit:
                self.__post_init__()
386
387
388
389
390
391

        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)]
392
            self.token_types = [[] for _ in range(self.n_seqs)]
393
            self.mrope_input_positions = None
394
395
396
397
398
399
            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

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
427
428
429
            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()
430
431
432
433
434

    def __init__(self,
                 runner: "GPUModelRunnerBase",
                 finished_requests_ids: Optional[List[str]] = None):
        super().__init__()
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
        # 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,
        ]

450
451
452
453
454
455
456
457
458
459
460
461
462
        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.decode_only = True

        # Attention metadata inputs.
463
464
465
466
        if self.attn_backend is not None:
            # spec decode (e.g. Medusa) does not have atten backend
            self.attn_metadata_builder = self.attn_backend.get_builder_cls()(
                weakref.proxy(self))
467
468
469
470
471
472
473
474
475
476
477

        # 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

478
479
480
481
482
483
484
485
486
487
488
    def prepare(self,
                finished_requests_ids: Optional[List[str]] = None) -> None:
        self.finished_requests_ids = finished_requests_ids

        # Intermediate data (data in CPU before going to GPU) for
        # the current sequence group.
        self.inter_data_list: List[
            ModelInputForGPUBuilder.InterDataForSeqGroup] = []

        self.attn_metadata_builder.prepare()

489
490
491
492
493
494
495
    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
496

497
498
        # Compute context length (the number of tokens that are
        # already computed) and sequence length (total number of tokens).
499

500
501
502
        seq_len = seq_data.get_len()
        if inter_data.is_prompt:
            context_len = seq_data.get_num_computed_tokens()
503
504
            seq_len = min(seq_len, context_len + token_chunk_size)
        elif self.runner.scheduler_config.is_multi_step or \
505
            self.runner.model_config.is_encoder_decoder:
506
            context_len = seq_len - 1
507
508
        else:
            context_len = seq_data.get_num_computed_tokens()
509
510

        # Compute tokens.
511
        tokens = seq_data.get_token_ids()[context_len:seq_len]
512
        token_types = seq_group_metadata.token_type_ids
513
514
515
516

        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
517
518
        inter_data.input_tokens[seq_idx].extend(tokens)
        inter_data.input_positions[seq_idx].extend(range(context_len, seq_len))
519
520
        inter_data.token_types[seq_idx].extend(
            token_types if token_types else [])
521
        inter_data.query_lens[seq_idx] = seq_len - context_len
522

523
524
525
526
527
528
529
530
531
532
533
        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,
                )

534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
    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
549
550
551
552
553
554
555
556
557

        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
558
559
560
        seq_group_metadata.seq_data[inter_data.seq_ids[
            seq_idx]].update_num_cached_tokens(prefix_cache_len)

561
562
563
564
565
566
567
568
569
570
571
572
573
        # 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
574
            inter_data.input_tokens[seq_idx] = inter_data.input_tokens[
575
                seq_idx][uncomputed_start:]
576
            inter_data.input_positions[seq_idx] = inter_data.input_positions[
577
                seq_idx][uncomputed_start:]
578
579
            inter_data.token_types[seq_idx] = inter_data.token_types[seq_idx][
                uncomputed_start:]
580
581
            context_len = prefix_cache_len

582
583
584
            inter_data.context_lens[seq_idx] = context_len
            inter_data.query_lens[
                seq_idx] = inter_data.seq_lens[seq_idx] - context_len
585
586
587
588
589
590
591
592
593
        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:]
594
595
            inter_data.token_types[seq_idx] = inter_data.token_types[seq_idx][
                -1:]
596
597
            inter_data.query_lens[seq_idx] = 1
            inter_data.context_lens[seq_idx] = inter_data.seq_lens[seq_idx] - 1
598
599
600
601
602
603
604
605
606
607
608
609
610
611

    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
612
613
614
615
616
617
            # 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
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634

        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)
635
636
637
638
639
640
641
        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([])
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673

    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."""
674
675
676
677
678
679
        # 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)))
680
681
682
        if not mm_data:
            return

683
684
685
686
687
688
689
690
691
        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
692
        inter_data.multi_modal_placeholder_maps = placeholder_maps
693

694
        # special processing for mrope position deltas.
695
        if self.runner.model_config.uses_mrope:
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
            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],
722
                        seq_len=inter_data.seq_lens[seq_idx],
723
724
725
726
727
728
                    )

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

729
    def add_seq_group(self, seq_group_metadata: SequenceGroupMetadata):
730
        """Add a sequence group to the builder."""
731
        seq_ids = seq_group_metadata.seq_data.keys()
732
733
734
735
736
737
738
        n_seqs = len(seq_ids)
        is_prompt = seq_group_metadata.is_prompt

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

739
740
        encoder_seq_len = 0

741
        if self.runner.model_config.is_encoder_decoder:
742
743
            encoder_seq_len = seq_group_metadata.encoder_seq_data.get_len()

744
        inter_data = self.init_cached_inter_data(
745
746
747
748
            request_id=seq_group_metadata.request_id,
            seq_ids=seq_ids,
            is_prompt=is_prompt,
            block_tables=seq_group_metadata.block_tables,
749
750
            computed_block_nums=seq_group_metadata.computed_block_nums,
            reinit=True,
751
752
            reinit_use_defaults=True,
            encoder_seq_len=encoder_seq_len)
753

754
        self.inter_data_list.append(inter_data)
755

756
757
758
759
760
        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)
761

762
763
    def _use_captured_graph(self,
                            batch_size: int,
764
                            decode_only: bool,
765
766
                            max_decode_seq_len: int,
                            max_encoder_seq_len: int = 0) -> bool:
767
        return (decode_only and not self.runner.model_config.enforce_eager
768
769
770
                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)
771

772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
    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:
788
            num_seqs (int): Number of sequences scheduled to run.
789
790
791
792
793
            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
794
                viability of using CUDA graphs.
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
        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

815
816
        graph_batch_size = self.runner.vllm_config.pad_for_cudagraph(
            batch_size)
817
818
819
        assert graph_batch_size >= batch_size
        return graph_batch_size - batch_size

820
    def build(self) -> ModelInputForGPU:
821
822
823
824
        """Finalize the builder intermediate data and
        create on-device tensors.
        """
        # Combine and flatten intermediate data.
825
        input_tokens = []
826
        token_types = []
827
828
829
        for inter_data in self.inter_data_list:
            for cur_input_tokens in inter_data.input_tokens:
                input_tokens.extend(cur_input_tokens)
830
831
            for cur_token_types in inter_data.token_types:
                token_types.extend(cur_token_types)
832

833
834
835
        if not input_tokens:
            # This may happen when all prefill requests hit
            # prefix caching and there is no decode request.
836
            return self.model_input_cls()
837

838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
        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)
859

860
        seq_lens = []
861
        query_lens = []
862
        max_decode_seq_len = 0
863
        max_encoder_seq_len = 0
864
865
        for inter_data in self.inter_data_list:
            seq_lens.extend(inter_data.seq_lens)
866
            query_lens.extend(inter_data.query_lens)
867
868
869
            if not inter_data.is_prompt:
                max_decode_seq_len = max(max_decode_seq_len,
                                         max(inter_data.seq_lens))
870
                if self.runner.model_config.is_encoder_decoder:
871
872
                    max_encoder_seq_len = max(max_encoder_seq_len,
                                              inter_data.encoder_seq_len)
873

874
875
876
877
878
879
        # 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
        }
880

881
882
        cuda_graph_pad_size = self._get_cuda_graph_pad_size(
            num_seqs=len(seq_lens),
883
            max_decode_seq_len=max_decode_seq_len,
884
            max_encoder_seq_len=max_encoder_seq_len)
885

886
887
888
889
890
891
        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
892
893

        # Tokens and positions.
894
895
        if cuda_graph_pad_size:
            input_tokens.extend(itertools.repeat(0, cuda_graph_pad_size))
896
897
898
899
        assert self.runner.device is not None
        input_tokens_tensor = async_tensor_h2d(input_tokens, torch.long,
                                               self.runner.device,
                                               self.runner.pin_memory)
900
901
902
903
904
905

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

906
907
908
909
910
911
912
913
914
915
916
917
918
919
        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)
920
        # Sequence and query lengths.
921
922
        if cuda_graph_pad_size:
            seq_lens.extend(itertools.repeat(1, cuda_graph_pad_size))
923
924
925

        # Attention metadata.
        attn_metadata = self.attn_metadata_builder.build(
926
            seq_lens, query_lens, cuda_graph_pad_size, batch_size)
927
928

        # LoRA data.
929
930
        lora_requests = set()
        lora_mapping = None
931
        if self.enable_lora:
932
933
934
935
936
937
            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
            ])
938
939
940
            if cuda_graph_pad_size:
                lora_index_mapping.extend(
                    itertools.repeat(0, cuda_graph_pad_size))
941
942
943
944
            lora_prompt_mapping = flatten_2d_lists([
                flatten_2d_lists(inter_data.lora_prompt_mapping)
                for inter_data in self.inter_data_list
            ])
945

946
            lora_mapping = LoRAMapping(
947
948
949
                **dict(index_mapping=lora_index_mapping,
                       prompt_mapping=lora_prompt_mapping,
                       is_prefill=not self.decode_only))
950
951

        # Prompt adapter data.
952
953
        prompt_adapter_requests: Set[PromptAdapterRequest] = set()
        prompt_adapter_mapping = None
954
        if self.enable_prompt_adapter:
955
956
957
958
959
960
961
            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
            ])
962
963
964
            if cuda_graph_pad_size:
                prompt_adapter_index_mapping.extend(
                    itertools.repeat(0, cuda_graph_pad_size))
965
966
967
968
            prompt_adapter_prompt_mapping = flatten_2d_lists([
                inter_data.prompt_adapter_prompt_mapping
                for inter_data in self.inter_data_list
            ])
969
            prompt_adapter_mapping = PromptAdapterMapping(
970
971
                prompt_adapter_index_mapping,
                prompt_adapter_prompt_mapping,
972
973
974
            )

        # Multi-modal data.
975
976
977
        multi_modal_kwargs_list = [
            data.multi_modal_kwargs for data in self.inter_data_list
            if data.multi_modal_kwargs is not None
978
        ]
979
        multi_modal_kwargs = MultiModalKwargs.batch(multi_modal_kwargs_list)
980
981
982
983

        return self.model_input_cls(
            input_tokens=input_tokens_tensor,
            input_positions=input_positions_tensor,
984
            token_types=token_types_tensor,
985
            attn_metadata=attn_metadata,
986
987
            seq_lens=seq_lens,
            query_lens=query_lens,
988
            lora_mapping=lora_mapping,
989
            lora_requests=lora_requests,
990
            multi_modal_kwargs=multi_modal_kwargs,
991
            request_ids_to_seq_ids=request_ids_to_seq_ids,
992
993
            finished_requests_ids=self.finished_requests_ids,
            prompt_adapter_mapping=prompt_adapter_mapping,
994
            prompt_adapter_requests=prompt_adapter_requests)
995
996


997
998
999
1000
1001
class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
    """
    Helper class for shared methods between GPU model runners.
    """
    _model_input_cls: Type[TModelInputForGPU]
1002
    _builder_cls: Type[ModelInputForGPUBuilder]
1003
    builder: ModelInputForGPUBuilder
1004
1005
1006

    def __init__(
        self,
1007
        vllm_config: VllmConfig,
1008
        kv_cache_dtype: Optional[str] = "auto",
1009
        is_driver_worker: bool = False,
1010
        return_hidden_states: bool = False,
1011
1012
        input_registry: InputRegistry = INPUT_REGISTRY,
        mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
1013
    ):
1014
1015
1016
1017
1018

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

1019
        self.is_driver_worker = is_driver_worker
1020
        self.return_hidden_states = return_hidden_states
1021

1022
        self.device = self.device_config.device
1023
        self.pin_memory = is_pin_memory_available()
1024

1025
1026
1027
1028
        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
1029
1030
        self.max_batchsize_to_capture = \
            self.vllm_config.compilation_config.max_capture_size
1031
1032
1033
1034

        self.graph_runners: List[Dict[int, CUDAGraphRunner]] = [
            {} for _ in range(self.parallel_config.pipeline_parallel_size)
        ]
1035
1036
        self.graph_memory_pool: Optional[Tuple[
            int, int]] = None  # Set during graph capture.
Mor Zusman's avatar
Mor Zusman committed
1037

1038
        self.has_inner_state = model_config.has_inner_state
Mor Zusman's avatar
Mor Zusman committed
1039

1040
1041
        self.in_profile_run = False

1042
        # When using CUDA graph, the input block tables must be padded to
1043
        # max_seq_len_to_capture. However, creating the block table in
1044
1045
1046
        # 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
1047
        # (max batch size to capture, max seq len to capture / block size).
1048
        self.graph_block_tables = np.zeros(
1049
            (self.max_batchsize_to_capture, self.get_max_block_per_batch()),
1050
            dtype=np.int32)
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061

        # 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)

1062
1063
1064
1065
1066
        self.attn_backend = get_attn_backend(
            self.model_config.get_head_size(),
            self.model_config.dtype,
            self.kv_cache_dtype,
            self.block_size,
1067
            self.model_config.is_attention_free,
1068
1069
1070
1071
1072
1073
        ) 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))
1074

1075
        # Multi-modal data support
1076
1077
1078
1079
        self.input_registry = input_registry
        self.mm_registry = mm_registry
        self.multi_modal_input_mapper = mm_registry \
            .create_input_mapper(model_config)
1080
        self.mm_registry.init_mm_limits_per_prompt(self.model_config)
1081

1082
        # Lazy initialization
1083
        self.model: nn.Module  # Set after load_model
1084
1085
        # Set after load_model.
        self.lora_manager: Optional[LRUCacheWorkerLoRAManager] = None
1086
        self.prompt_adapter_manager: LRUCacheWorkerPromptAdapterManager = None
1087

1088
1089
1090
        set_cpu_offload_max_bytes(
            int(self.cache_config.cpu_offload_gb * 1024**3))

1091
1092
        # Used to cache python objects
        self.inter_data_cache: Dict[int, PyObjectCache] = {}
1093
1094
1095
1096
1097
1098
1099

        # 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.
1100
        self.sampling_metadata_cache: SamplingMetadataCache = \
1101
1102
              SamplingMetadataCache() \
                if self.parallel_config.pipeline_parallel_size == 1 else None
1103

1104
1105
1106
1107
        if hasattr(self, "_builder_cls"):
            # multi-step model runner does not have `_builder_cls`
            self.builder = self._builder_cls(weakref.proxy(self))

1108
    def load_model(self) -> None:
1109
        logger.info("Starting to load model %s...", self.model_config.model)
1110
        with DeviceMemoryProfiler() as m:
1111
            self.model = get_model(vllm_config=self.vllm_config)
1112
1113

        self.model_memory_usage = m.consumed_memory
1114
1115
        logger.info("Loading model weights took %.4f GB",
                    self.model_memory_usage / float(2**30))
1116
1117

        if self.lora_config:
1118
            assert supports_lora(
1119
                self.model
1120
            ), f"{self.model.__class__.__name__} does not support LoRA yet."
1121

1122
1123
1124
            if supports_multimodal(self.model):
                logger.warning("Regarding multimodal models, vLLM currently "
                               "only supports adding LoRA to language model.")
1125
1126
1127
1128
1129
1130
1131
            # 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)
1132

1133
1134
            self.lora_manager = LRUCacheWorkerLoRAManager(
                self.scheduler_config.max_num_seqs,
1135
1136
1137
1138
1139
1140
                self.scheduler_config.max_num_batched_tokens,
                self.vocab_size,
                self.lora_config,
                self.device,
                self.model.embedding_modules,
                self.model.embedding_padding_modules,
1141
                max_position_embeddings=max_pos_embeddings,
1142
            )
1143
            self.model = self.lora_manager.create_lora_manager(self.model)
1144

1145
1146
1147
1148
1149
1150
1151
1152
1153
        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))

1154
1155
        if self.kv_cache_dtype == "fp8" and (current_platform.is_rocm()
                                             or current_platform.is_cuda()):
1156
1157
1158
            # Currently only ROCm accepts kv-cache scaling factors
            # via quantization_param_path and this will be deprecated
            # in the future.
1159
1160
            if self.model_config.quantization_param_path is not None:
                if callable(getattr(self.model, "load_kv_cache_scales", None)):
1161
1162
1163
1164
1165
1166
                    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)
1167
1168
                    self.model.load_kv_cache_scales(
                        self.model_config.quantization_param_path)
1169
1170
                    logger.info("Loaded KV cache scaling factors from %s",
                                self.model_config.quantization_param_path)
1171
                else:
1172
1173
1174
1175
                    raise RuntimeError(
                        "Using FP8 KV cache and scaling factors provided but "
                        "model %s does not support loading scaling factors.",
                        self.model.__class__)
1176
            else:
1177
1178
1179
1180
                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!")
1181

1182
1183
        if self.vllm_config.compilation_config.level ==\
            CompilationLevel.DYNAMO_AS_IS and supports_dynamo():
1184
1185
            backend = self.vllm_config.compilation_config.init_backend(
                self.vllm_config)
1186
1187
1188
            self.model = torch.compile(
                self.model,
                fullgraph=envs.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE,
1189
                backend=backend)
1190

1191
1192
1193
    def get_model(self) -> nn.Module:
        return self.model

1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
    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,
        )

1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
    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,
        )

1218
1219
    def get_max_block_per_batch(self) -> int:
        block_size = self.block_size
1220
        return (self.max_seq_len_to_capture + block_size - 1) // block_size
1221

1222
    def _prepare_model_input_tensors(
1223
1224
        self,
        seq_group_metadata_list: List[SequenceGroupMetadata],
Mor Zusman's avatar
Mor Zusman committed
1225
        finished_requests_ids: Optional[List[str]] = None
1226
1227
1228
1229
    ) -> 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.
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240

        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.
        """
1241
        self.builder.prepare(finished_requests_ids)
1242
        for seq_group_metadata in seq_group_metadata_list:
1243
            self.builder.add_seq_group(seq_group_metadata)
1244

1245
        self.builder.reset_cached_inter_data()
1246

1247
        return self.builder.build()  # type: ignore
1248

1249
1250
1251
1252
1253
1254
1255
1256
    @contextmanager
    def set_in_profile_run(self):
        self.in_profile_run = True
        try:
            yield
        finally:
            self.in_profile_run = False

1257
1258
    @torch.inference_mode()
    def profile_run(self) -> None:
1259
1260
1261
1262
1263
1264
1265
1266
        max_num_batched_tokens = \
            self.scheduler_config.max_num_batched_tokens
        max_num_seqs = self.scheduler_config.max_num_seqs
        self._dummy_run(max_num_batched_tokens, max_num_seqs)

    def _dummy_run(self,
                   max_num_batched_tokens: int,
                   max_num_seqs: int = 1) -> None:
1267
1268
1269
1270
        with self.set_in_profile_run():
            # Enable top-k sampling to reflect the accurate memory usage.
            sampling_params = \
                SamplingParams(top_p=0.99, top_k=self.vocab_size - 1)
1271

1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
            # This represents the maximum number of different requests
            # that will have unique loras, an therefore the max amount of memory
            # consumption create dummy lora request copies from the lora request
            # passed in, which contains a lora from the lora warmup path.
            dummy_lora_requests: List[LoRARequest] = []
            dummy_lora_requests_per_seq: List[LoRARequest] = []
            if self.lora_config:
                assert self.lora_manager is not None
                with self.lora_manager.dummy_lora_cache():
                    for idx in range(self.lora_config.max_loras):
                        lora_id = idx + 1
                        dummy_lora_request = LoRARequest(
                            lora_name=f"warmup_{lora_id}",
                            lora_int_id=lora_id,
                            lora_path="/not/a/real/path",
                        )
                        self.lora_manager.add_dummy_lora(dummy_lora_request,
                                                         rank=LORA_WARMUP_RANK)
                        dummy_lora_requests.append(dummy_lora_request)
                    dummy_lora_requests_per_seq = [
                        dummy_lora_requests[idx % len(dummy_lora_requests)]
                        for idx in range(max_num_seqs)
                    ]

            # Profile memory usage with max_num_sequences sequences and the
            # total number of tokens equal to max_num_batched_tokens.
            seqs: List[SequenceGroupMetadata] = []
            # Additional GPU memory may be needed for multi-modal encoding,
            # which needs to be accounted for when calculating the GPU blocks
            # for vLLM blocker manager.
            # To exercise the worst scenario for GPU memory consumption,
            # the number of seqs (batch_size) is chosen to maximize the number
            # of images processed.

            max_mm_tokens = self.mm_registry.get_max_multimodal_tokens(
                self.model_config)
            if max_mm_tokens > 0:
                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

            batch_size = 0
            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))
                batch_size += seq_len

                dummy_data = self.input_registry \
                    .dummy_data_for_profiling(self.model_config,
                                            seq_len,
                                            self.mm_registry)

                seq = SequenceGroupMetadata(
                    request_id=str(group_id),
                    is_prompt=True,
                    seq_data={group_id: dummy_data.seq_data},
                    sampling_params=sampling_params,
                    block_tables=None,
                    lora_request=dummy_lora_requests_per_seq[group_id]
                    if dummy_lora_requests_per_seq else None,
                    multi_modal_data=dummy_data.multi_modal_data,
                    multi_modal_placeholders=dummy_data.
                    multi_modal_placeholders,
                )
                seqs.append(seq)

            # Run the model with the dummy inputs.
            num_layers = self.model_config.get_num_layers(self.parallel_config)
            # 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).
            # it is important to create tensors inside the loop, rather than
            # multiplying the list, to avoid Dynamo from treating them as
            # tensor aliasing.
            kv_caches = [
                torch.tensor([], dtype=torch.float32, device=self.device)
                for _ in range(num_layers)
            ]
            finished_requests_ids = [seq.request_id for seq in seqs]
            model_input = self.prepare_model_input(
                seqs, finished_requests_ids=finished_requests_ids)
            intermediate_tensors = None
            if not get_pp_group().is_first_rank:
                intermediate_tensors = \
                    self.model.make_empty_intermediate_tensors(
                    batch_size=batch_size,
                    dtype=self.model_config.dtype,
                    device=self.device)

            self.execute_model(model_input, kv_caches, intermediate_tensors)
            torch.cuda.synchronize()
            return
1372

1373
    def remove_all_loras(self):
1374
1375
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
1376
        self.lora_manager.remove_all_adapters()
1377

1378
    def set_active_loras(self, lora_requests: Set[LoRARequest],
1379
1380
1381
                         lora_mapping: LoRAMapping) -> None:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
1382
        self.lora_manager.set_active_adapters(lora_requests, lora_mapping)
1383
1384
1385
1386

    def add_lora(self, lora_request: LoRARequest) -> bool:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
1387
        return self.lora_manager.add_adapter(lora_request)
1388
1389
1390
1391

    def remove_lora(self, lora_id: int) -> bool:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
1392
        return self.lora_manager.remove_adapter(lora_id)
1393
1394
1395
1396

    def pin_lora(self, lora_id: int) -> bool:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
1397
        return self.lora_manager.pin_adapter(lora_id)
1398
1399
1400
1401

    def list_loras(self) -> Set[int]:
        if not self.lora_manager:
            raise RuntimeError("LoRA is not enabled.")
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
        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()
1437

1438
    @torch.inference_mode()
1439
    def capture_model(self, kv_caches: List[List[torch.Tensor]]) -> None:
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
        """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.
        """
1452
        assert not self.model_config.enforce_eager
1453
        logger.info("Capturing cudagraphs for decoding. This may lead to "
1454
1455
                    "unexpected consequences if the model is not static. To "
                    "run the model in eager mode, set 'enforce_eager=True' or "
1456
1457
                    "use '--enforce-eager' in the CLI. "
                    "If out-of-memory error occurs during cudagraph capture,"
1458
1459
1460
                    " consider decreasing `gpu_memory_utilization` or "
                    "switching to eager mode. You can also reduce the "
                    "`max_num_seqs` as needed to decrease memory usage.")
1461
        start_time = time.perf_counter()
1462
        start_free_gpu_memory = torch.cuda.mem_get_info()[0]
1463
1464

        # Prepare dummy inputs. These will be reused for all batch sizes.
1465
        max_batch_size = self.max_batchsize_to_capture
1466
1467
1468
1469
1470
1471
        input_tokens = torch.zeros(max_batch_size,
                                   dtype=torch.long,
                                   device=self.device)
        input_positions = torch.zeros(max_batch_size,
                                      dtype=torch.long,
                                      device=self.device)
1472
        if self.model_config.uses_mrope:
1473
1474
            input_positions = torch.tile(input_positions,
                                         (3, 1)).cuda(device=self.device)
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
        # 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)

1486
1487
1488
1489
1490
1491
        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)
1492

1493
1494
        with self.attn_state.graph_capture(max_batch_size), graph_capture(
                self.device) as graph_capture_context:
1495
1496
            # NOTE: Capturing the largest batch size first may help reduce the
            # memory usage of CUDA graph.
1497
1498
            for virtual_engine in range(
                    self.parallel_config.pipeline_parallel_size):
1499
                # Only rank 0 should print progress bar during capture
1500
1501
1502
1503
1504
1505
1506
1507
                cudagraph_capture_sizes = (tqdm(
                    self.vllm_config.compilation_config.
                    cudagraph_capture_sizes,
                    desc="Capturing CUDA graph shapes",
                ) if get_tensor_model_parallel_rank() == 0 else
                                           self.vllm_config.compilation_config.
                                           cudagraph_capture_sizes)
                for batch_size in cudagraph_capture_sizes:
1508
1509
                    attn_metadata = (
                        self.attn_state.graph_capture_get_metadata_for_batch(
1510
1511
                            batch_size,
                            is_encoder_decoder_model=self.model_config.
1512
                            is_encoder_decoder))
1513
1514
1515

                    if self.lora_config:
                        lora_mapping = LoRAMapping(
1516
1517
1518
                            **dict(index_mapping=[0] * batch_size,
                                   prompt_mapping=[0] * batch_size,
                                   is_prefill=False))
1519
1520
                        self.set_active_loras(set(), lora_mapping)

1521
1522
1523
1524
1525
1526
1527
                    if self.prompt_adapter_config:
                        prompt_adapter_mapping = PromptAdapterMapping(
                            [-1] * batch_size,
                            [-1] * batch_size,
                        )
                        self.set_active_prompt_adapters(
                            set(), prompt_adapter_mapping)
1528
                    graph_runner = CUDAGraphRunner(
1529
                        self.model, self.attn_backend.get_name(),
1530
                        self.attn_state.graph_clone(batch_size),
1531
                        self.model_config.is_encoder_decoder)
1532

Mor Zusman's avatar
Mor Zusman committed
1533
1534
                    capture_inputs = {
                        "input_ids":
1535
                        input_tokens[:batch_size],
Mor Zusman's avatar
Mor Zusman committed
1536
                        "positions":
1537
                        input_positions[..., :batch_size],
Mor Zusman's avatar
Mor Zusman committed
1538
                        "intermediate_inputs":
1539
1540
                        intermediate_inputs[:batch_size]
                        if intermediate_inputs is not None else None,
Mor Zusman's avatar
Mor Zusman committed
1541
                        "kv_caches":
1542
                        kv_caches[virtual_engine],
Mor Zusman's avatar
Mor Zusman committed
1543
                        "attn_metadata":
1544
                        attn_metadata,
Mor Zusman's avatar
Mor Zusman committed
1545
1546
1547
1548
1549
                        "memory_pool":
                        self.graph_memory_pool,
                        "stream":
                        graph_capture_context.stream
                    }
1550
1551
1552
1553
1554
                    if previous_hidden_states is not None:
                        capture_inputs[
                            "previous_hidden_states"] = previous_hidden_states[:
                                                                               batch_size]

1555
                    if self.has_inner_state:
Mor Zusman's avatar
Mor Zusman committed
1556
1557
1558
1559
1560
1561
                        # 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)
                        })
1562
                    if self.model_config.is_encoder_decoder:
1563
1564
1565
1566
1567
                        # add the additional inputs to capture for
                        # encoder-decoder models.
                        self._update_inputs_to_capture_for_enc_dec_model(
                            capture_inputs)

1568
1569
                    with set_forward_context(attn_metadata, self.vllm_config,
                                             virtual_engine):
1570
                        graph_runner.capture(**capture_inputs)
1571
1572
1573
                    self.graph_memory_pool = graph_runner.graph.pool()
                    self.graph_runners[virtual_engine][batch_size] = (
                        graph_runner)
1574
1575

        end_time = time.perf_counter()
1576
        end_free_gpu_memory = torch.cuda.mem_get_info()[0]
1577
        elapsed_time = end_time - start_time
1578
        cuda_graph_size = start_free_gpu_memory - end_free_gpu_memory
1579
        # This usually takes < 10 seconds.
1580
1581
        logger.info("Graph capturing finished in %.0f secs, took %.2f GiB",
                    elapsed_time, cuda_graph_size / GiB_bytes)
1582

1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
    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.
1596
1597
1598
1599
1600
1601
        capture_inputs["encoder_input_ids"] = torch.tensor([],
                                                           dtype=torch.long,
                                                           device=self.device)
        capture_inputs["encoder_positions"] = torch.tensor([],
                                                           dtype=torch.long,
                                                           device=self.device)
1602

1603
1604
1605
1606
    @property
    def vocab_size(self) -> int:
        return self.model_config.get_vocab_size()

1607

1608
1609
1610
1611
1612
1613
class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
    """
    GPU model runner with sampling step.
    """
    _model_input_cls: Type[ModelInputForGPUWithSamplingMetadata] = (
        ModelInputForGPUWithSamplingMetadata)
1614
    _builder_cls: Type[ModelInputForGPUBuilder] = ModelInputForGPUBuilder
1615
1616
1617
1618
1619

    def make_model_input_from_broadcasted_tensor_dict(
        self,
        tensor_dict: Dict[str, Any],
    ) -> ModelInputForGPUWithSamplingMetadata:
1620
        model_input = \
1621
1622
1623
            ModelInputForGPUWithSamplingMetadata.from_broadcasted_tensor_dict(
                tensor_dict,
                attn_backend=self.attn_backend,
1624
1625
            )
        return model_input
1626
1627
1628
1629

    def prepare_model_input(
        self,
        seq_group_metadata_list: List[SequenceGroupMetadata],
1630
        virtual_engine: int = 0,
1631
        finished_requests_ids: Optional[List[str]] = None,
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
    ) -> 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
1647
            seq_group_metadata_list, finished_requests_ids)
1648
1649
1650
1651
1652
1653
        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,
1654
                generators, self.sampling_metadata_cache)
1655
1656
        else:
            sampling_metadata = None
1657
1658
1659
1660
        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,
1661
1662
                                   is_prompt=is_prompt,
                                   virtual_engine=virtual_engine)
1663
1664

    @torch.inference_mode()
1665
    @dump_input_when_exception(exclude_args=[0], exclude_kwargs=["self"])
1666
1667
1668
1669
    def execute_model(
        self,
        model_input: ModelInputForGPUWithSamplingMetadata,
        kv_caches: List[torch.Tensor],
1670
        intermediate_tensors: Optional[IntermediateTensors] = None,
1671
        num_steps: int = 1,
1672
    ) -> Optional[Union[List[SamplerOutput], IntermediateTensors]]:
1673
1674
1675
        if num_steps > 1:
            raise ValueError("num_steps > 1 is not supported in ModelRunner")

1676
1677
1678
1679
1680
1681
        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)

1682
1683
1684
1685
1686
1687
1688
        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)

1689
        self.attn_state.begin_forward(model_input)
1690

1691
1692
1693
1694
        # 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
1695
1696
1697
        # TODO(andoorve): We can remove this once all
        # virtual engines share the same kv cache.
        virtual_engine = model_input.virtual_engine
1698
1699
1700
        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]
1701
1702
            model_executable = self.graph_runners[virtual_engine][
                graph_batch_size]
1703
1704
1705
        else:
            model_executable = self.model

1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
        # 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
                )

1724
        multi_modal_kwargs = model_input.multi_modal_kwargs or {}
Mor Zusman's avatar
Mor Zusman committed
1725
1726
1727
        seqlen_agnostic_kwargs = {
            "finished_requests_ids": model_input.finished_requests_ids,
            "request_ids_to_seq_ids": model_input.request_ids_to_seq_ids,
1728
        } if self.has_inner_state else {}
1729
1730
1731
1732
1733
1734
        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()

1735
1736
        if not bypass_model_exec:
            with set_forward_context(model_input.attn_metadata,
1737
                                     self.vllm_config, virtual_engine):
1738
1739
1740
1741
1742
1743
1744
1745
1746
                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)
1747

1748
1749
1750
1751
        if (self.observability_config is not None
                and self.observability_config.collect_model_forward_time):
            model_forward_end.record()

1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
        # 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,
            )

1765
1766
        # Compute the logits in the last pipeline stage.
        if not get_pp_group().is_last_rank:
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
            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))
1782
1783
1784
            return hidden_or_intermediate_states

        logits = self.model.compute_logits(hidden_or_intermediate_states,
1785
1786
1787
                                           model_input.sampling_metadata)

        if not self.is_driver_worker:
1788
            return []
1789

1790
1791
        if model_input.async_callback is not None:
            model_input.async_callback()
1792

1793
1794
1795
1796
1797
        # Sample the next token.
        output: SamplerOutput = self.model.sample(
            logits=logits,
            sampling_metadata=model_input.sampling_metadata,
        )
1798
1799
1800
1801
1802
1803
        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)
1804
1805
1806
1807
            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()
1808
1809
1810
1811
            # 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.
1812
1813
            output.model_forward_time = (orig_model_forward_time +
                                         model_forward_time)
1814
1815
1816

        if self.return_hidden_states:
            # we only need to pass hidden states of most recent token
1817
1818
            assert model_input.sampling_metadata is not None
            indices = model_input.sampling_metadata.selected_token_indices
1819
            if model_input.is_prompt:
1820
1821
                hidden_states = hidden_or_intermediate_states.index_select(
                    0, indices)
1822
                output.prefill_hidden_states = hidden_or_intermediate_states
1823
            elif decode_meta.use_cuda_graph:
1824
1825
1826
                hidden_states = hidden_or_intermediate_states[:len(indices)]
            else:
                hidden_states = hidden_or_intermediate_states
1827

1828
1829
            output.hidden_states = hidden_states

1830
        return [output]
1831

1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
    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
1844
1845
1846
        if self.vllm_config.kv_transfer_config is None:
            return False

1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
        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
1869
1870
1871
        if self.vllm_config.kv_transfer_config is None:
            return False

1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
        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

1882

1883
1884
1885
# NOTE: this is nn.Module so the profiler can properly capture/group
#  kernels calls made within the graph
class CUDAGraphRunner(nn.Module):
1886

1887
    def __init__(self, model: nn.Module, backend_name: str,
1888
                 attn_state: AttentionState, is_encoder_decoder_model: bool):
1889
        super().__init__()
1890
        self.model = model
1891
        self.backend_name = backend_name
1892
        self.attn_state = attn_state
1893

1894
1895
1896
        self.input_buffers: Dict[str, torch.Tensor] = {}
        self.output_buffers: Dict[str, torch.Tensor] = {}

1897
        self._graph: Optional[torch.cuda.CUDAGraph] = None
1898
        self._is_encoder_decoder_model = is_encoder_decoder_model
1899
1900
1901
1902
1903
1904

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

1905
1906
1907
1908
    def capture(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1909
        intermediate_inputs: Optional[IntermediateTensors],
1910
1911
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
1912
1913
        memory_pool: Optional[Tuple[int, int]],
        stream: torch.cuda.Stream,
1914
        **kwargs,
1915
    ):
1916
        assert self._graph is None
1917
        # Run the model a few times without capturing the graph.
1918
1919
        # This is to make sure that the captured graph does not include the
        # kernel launches for initial benchmarking (e.g., Triton autotune).
1920
        # Note one iteration is not enough for torch.compile
1921
1922
        for _ in range(_NUM_WARMUP_ITERS):
            self.model(
1923
1924
1925
1926
1927
                input_ids=input_ids,
                positions=positions,
                kv_caches=kv_caches,
                attn_metadata=attn_metadata,
                intermediate_tensors=intermediate_inputs,
1928
1929
                **kwargs,
            )
1930
1931
        # Wait for the warm up operations to finish before proceeding with
        # Graph Capture.
1932
1933
1934
1935
        torch.cuda.synchronize()
        # Capture the graph.
        self._graph = torch.cuda.CUDAGraph()
        with torch.cuda.graph(self._graph, pool=memory_pool, stream=stream):
1936
            output_hidden_or_intermediate_states = self.model(
1937
1938
1939
1940
1941
                input_ids=input_ids,
                positions=positions,
                kv_caches=kv_caches,
                attn_metadata=attn_metadata,
                intermediate_tensors=intermediate_inputs,
1942
                **kwargs,
1943
            )
1944
1945
1946

            if isinstance(output_hidden_or_intermediate_states, torch.Tensor):
                hidden_or_intermediate_states = weak_ref_tensor(
1947
                    output_hidden_or_intermediate_states)
1948
1949
1950
1951
1952
1953
1954
1955
            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()
                    })
1956
1957

            del output_hidden_or_intermediate_states
1958
            # make sure `output_hidden_or_intermediate_states` is deleted
1959
1960
            # in the graph's memory pool
            gc.collect()
1961
1962
1963
        torch.cuda.synchronize()

        # Save the input and output buffers.
1964
        self.input_buffers = {
1965
1966
1967
1968
1969
1970
1971
1972
            "input_ids":
            input_ids,
            "positions":
            positions,
            "kv_caches":
            kv_caches,
            **self.attn_state.get_graph_input_buffers(
                attn_metadata, self._is_encoder_decoder_model),
1973
1974
            **kwargs,
        }
1975
1976
1977
1978
1979
1980
1981
1982
        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
1983
1984
1985
1986
1987

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1988
1989
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
1990
        intermediate_tensors: Optional[IntermediateTensors],
1991
        **kwargs,
1992
1993
1994
1995
1996
    ) -> 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.
1997
1998
        self.input_buffers["input_ids"].copy_(input_ids, non_blocking=True)
        self.input_buffers["positions"].copy_(positions, non_blocking=True)
1999

2000
        if self.backend_name != "NO_ATTENTION":
2001
2002
2003
            self.input_buffers["slot_mapping"].copy_(
                attn_metadata.slot_mapping, non_blocking=True)

2004
2005
        self.attn_state.prepare_graph_input_buffers(
            self.input_buffers, attn_metadata, self._is_encoder_decoder_model)
2006

Mor Zusman's avatar
Mor Zusman committed
2007
2008
2009
        if "seqlen_agnostic_capture_inputs" in self.input_buffers:
            self.model.copy_inputs_before_cuda_graphs(self.input_buffers,
                                                      **kwargs)
2010
2011
2012
2013
2014

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

2015
2016
        if intermediate_tensors is not None:
            for key in intermediate_tensors.tensors:
2017
                if key != "model_execute_time" and key != "model_forward_time":
2018
2019
                    self.input_buffers[key].copy_(intermediate_tensors[key],
                                                  non_blocking=True)
2020
2021
2022
2023
2024
2025
        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)

2026
2027
2028
        # Run the graph.
        self.graph.replay()
        # Return the output tensor.
2029
2030
2031
2032
        if get_pp_group().is_last_rank:
            return self.output_buffers["hidden_states"]

        return self.output_buffers